{ "model_id": "m4", "model_tag": "mistral-small3.2:24b", "arm": "one", "box_class": "gpu-5090-laptop-24g", "mode": "context-ladder", "arm_cards": [ "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff" ], "target_uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "started_utc": "2026-09-22T00:25:59Z", "prompt_sha256": "90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c", "num_predict": 256, "ladder_note": "The frozen prompt is about 537 tokens. Climbing this ladder measures the cost of RESERVING a context window, NOT the cost of filling one: a 128k window with 537 tokens in it decodes at nearly the speed of an 8k window with 537 tokens in it. What changes is the VRAM the reservation costs, and therefore what fits. Any tok/s difference across levels is a second-order effect of the allocation and is not a long-prompt number. The FILLED-WINDOW arm in this same directory is the one that answers the long-prompt question, and the two must never be quoted as if they were the same measurement.", "fit_rule_note": "ONE RULE ON THIS BENCH, and it is the strict one, inherited unchanged from the one-3090 bench: a context rung FITS when `/api/ps` reports `size_vram == size` -- every byte of weights and KV on the card, nothing in host RAM -- and the runner's own resident memory stays inside the allowance below; a rung that spills is printed as `does not fit` with its per-card memory and the ladder stops there. The 10 and 12 GB legs of this ladder run a second, looser rule (`any-vram`) because NOTHING in the bank fits those boards. This board is 24 GB and both bank models fit it whole, so that rule is not in play here and inheriting it would have let a spilled figure be published in a column headed by a whole one. The `cpu-box` arm keeps its own `cpu-only` rule and is registered rather than run.", "ladder_stop_note": "On this 24 GB board the ladder climbs until a rung SPILLS and stops there, because the spill IS the ceiling: the headline is the largest context window the board held WHOLE. That is the same stop the one-3090 bench used (its own headlines: gemma4:26b at 131,072 and mistral-small3.2:24b at 65,536), and using the same stop is what lets this board's headline be compared with those rather than explained against them. The HOST-side stop the 10 and 12 GB legs use does not apply here and is not inherited: it exists for boards on which every rung spills by construction, where a VRAM stop would fire at the first rung and measure nothing.", "fit_rule_by_arm": { "one": "whole", "cpu-box": "cpu-only" }, "cap_axis_note": "this board's entire power envelope -- 95 W default, 175 W maximum, both READ off the card -- sits BELOW the lowest cap column the 24 GB comparison tables have, which is 250 W. A row from this bench therefore fills ZERO cells in any cap-keyed table, however well the bench runs, and no reducer may place one there. The comparable form is a table whose columns are READINGS with the cap printed beside every figure. The finding this bench exists for is not 'the laptop is slower'; it is what a part allowed a fifth of the power does with the same model, the same prompt and the same instrument.", "cap_dynamic_boost_note": "dynamic-boost -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw. Not a rung: its cap cell is the RANGE plus the mean enforced limit over the run, derived from the 2 Hz trace, and never a single number.", "pcie_note": "This card negotiated PCIe x16 of a x16-capable generation-3 link -- the same x16 slot the 3090 used, and the slot on which the pair bench measured a x16 + x4 asymmetry. With one card there is no per-token traffic BETWEEN cards, so the link is not in the decode path the way it was for the split pair; it still carries the model in and the tokens out. Every run records `pcie.link.width.current` UNDER LOAD, because an idle card drops its link to save power and a width read at rest would flatter the result.", "thermal_layout_note": "ONE GeForce RTX 3090 Ti 24 GB in a consumer desktop, alone on the board's x16 slot -- the same box, the same case and the same airflow in which one GeForce RTX 3090 24 GB was measured earlier the same day and two GeForce RTX 3080 10 GB cards the night before. A temperature here is a reading of THAT arrangement. Against the single 3090 the arrangement is the same one, which is what makes the two cards' thermal rows comparable; against the PAIR it is not, because with one card there is no second board warming the air or blocking a face. The card is traced on every run.", "card_labels": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": "soldered (mobile; no socket \u2014 link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c" }, "card_identity": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": null, "bus_id": "00000000:01:00.0", "pci_sub_device_id": null, "vendor": null, "board_per_operator": null, "vbios": null, "serial": null, "pcie_link_width_max": null, "pcie_link_width_negotiated": null, "memory_total_mib": null, "memory_bus_width_bits": null, "clocks_max_sm_mhz": null, "clocks_max_memory_mhz": null, "power_default_limit_w": null, "power_min_limit_w": null, "power_max_limit_w": null, "power_limit_at_bench_start_w": null, "note": "the single NVIDIA GeForce RTX 5090 Laptop GPU 24 GB SOLDERED to this machine's board at 00000000:01:00.0. There is no seat, no partner and no card to swap. \u26a0 THE TRAP ON THIS BOX IS NOT A STALE BOOT UNIT, IT IS A RUNNING DAEMON: nvidia-powerd.service (NVIDIA Dynamic Boost) floats this board's power limit between its default and its maximum against the CPU's draw, continuously and without asking, and it has been running since 2026-09-08. The lesson every leg of this ladder carries \u2014 a cap read once at the start of a night is not a cap \u2014 is true here in its strongest form: such a reading is a sample of a moving signal. Every arm reads the limit back in the same call as its own data, every stage reads it again at its close, and a stage that finds it moved STOPS rather than relabelling its file. The second trap is a CLOCK LOCK: ai-perf.service runs `nvidia-smi -lgc 1200,2550` at boot on this box and the cards that produced the rows this bench compares against ran unlocked, so a floored clock changes what a power cap means and every record carries clocks.sm min and max so the confound is visible rather than implied." } }, "memory_temp_support": { "checked_utc": "2026-09-22T00:25:59Z", "paths": { "nvidia-smi --query-gpu=temperature.memory": { "raw": "0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, N/A", "readable": false }, "nvidia-smi -q -d TEMPERATURE": { "lines": [ "GPU Target Temperature : 87 C", "Memory Current Temp : N/A", "Memory Max Operating T.Limit Temp : N/A" ], "readable": false }, "NVML NVML_FI_DEV_MEMORY_TEMP": { "available": true, "field_id": 82, "cards": { "0": { "call_rc": 0, "field_rc": 3, "supported": false, "not_supported": true, "value_c": null } } } }, "readable": false, "verdict": "the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead." }, "runner_rss_allowance_gib": 4.0, "num_gpu_option": null, "num_gpu_policy": "layers left to ollama's own planner (`auto`)", "num_gpu_note": "ollama decides for itself how many of a model's layers to put on the card. On the pair of 3080s that decision was measurably conservative -- gemma4:26b loaded 75.2% into VRAM and spilled 4.25 GiB to host RAM while about 5 GiB of the two cards' 20 GiB sat unused -- so the pair bench reported `auto` and a forced layer count side by side. On ONE 24 GB card there is no placement decision to make: the expected reading is 100% on the card with no spill. This harness records `options.num_gpu` on every run as `num_gpu_option` and reads the planner's actual placement back from /api/ps, so a spill on a card with room is visible as a finding rather than absorbed into a tok/s figure. A value at or above the model's layer count means every layer. ON THIS CARD THE QUESTION INVERTS. The pair bench forced `num_gpu` UP to recover cards the planner had left half empty; on a 12 GB board the model cannot fit however high the count goes, so forcing it up is not a recovery -- it is either ignored or an out-of-memory refusal, and either is a result. What is worth measuring is whether the planner leaves VRAM on the table on a card it CANNOT fill: so the probe BRACKETS the planner's own choice rather than jumping to the layer count. The planner's num_gpu is read back from the server, the model's own block count is read from /api/show (never typed -- the pair bench knew gemma4:26b had 31 layers because it READ it, and this bench has no receipt for the dense model's), and the probe walks UPWARD from the planner's choice in the steps PREREG SS5.4 registers, stopping at the FIRST refusal and recording it. An out-of-memory answer is written into the result file as that rung's outcome, never as a discarded run.", "instance_env": { "kind": "instance-env-receipt", "box_class": "gpu-5090-laptop-24g", "target_uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "written_utc": "2026-09-22T00:23:09Z", "unit": "bench-5090laptop-one", "shape": "one", "port": 11470, "base_url": "http://127.0.0.1:11470", "cuda_visible_devices": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "OLLAMA_FLASH_ATTENTION": "1", "OLLAMA_KV_CACHE_TYPE": "q8_0", "OLLAMA_CONTEXT_LENGTH": 32768, "OLLAMA_": 1, "OLLAMA_MAX_LOADED_MODELS": 1, "OLLAMA_KEEP_ALIVE": "0 (the server default; every bench request sends its own keep_alive, which wins, and every arm unloads on the way out)", "models_dir": "/usr/share/ollama/.ollama/models", "models_dir_writable_by_this_user": false, "api_version": { "version": "0.32.13" }, "ollama_binary": "/workshop/bench-laptop-5090-2026-09-21/ollama-0.32.13/bin/ollama", "ollama_client_version": "0.32.13", "ollama_version_pin": "0.32.13", "disclosed_tenants": "nomic-embed-text:latest on the SYSTEM ollama (:, OLLAMA_KEEP_ALIVE=-1, ~323 MB VRAM) \u2014 disclosed, never unloaded", "power_and_clocks_at_start_csv": "[N/A], 150.00 W, 95.00 W, 175.00 W, 1192 MHz, 810 MHz, 3090 MHz", "power_and_clocks_at_start_fields": "power.limit,enforced.power.limit,power.default_limit,power.max_limit,clocks.sm,clocks.mem,clocks.max.sm", "nvidia_powerd_at_start": "active", "clock_lock_unit_at_start": "active", "clock_lock_unit_name": "ai-perf.service", "clock_lock_range_declared": "1200,2550", "clock_lock_unit_state_is_not_evidence": "ai-perf.service is Type=oneshot and reads 'active (exited)' for the whole uptime whatever happens to the card afterwards. The lock's real state is read from the CARD (clocks.sm against the lock's floor) and is recorded in this leg's clock-lock receipt.", "power_limits_at_start_w": "0, [N/A], 150.00 W;", "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 8;" }, "ups_before_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-22T00:25:59Z", "available": false, "error": "no upsc on this box" }, "levels": [ { "num_ctx": 4096, "started_utc": "2026-09-22T00:25:59Z", "loads": true, "size_total": 14602993663, "size_vram": 14602993663, "context_length_loaded": 4096, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 14799.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 274.0 }, "runner_peak_rss_gib": 0.82, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1818421, "rss_kb": 862508, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27138 13472 4450 30129 35560\nSwap: 18431 18431 0", "kv_footprint_bytes": 0, "kv_footprint_gib": 0.0, "runs": [ { "ttft_ms": 230.08, "first_token_was_thinking": false, "wall_s": 5.9568, "eval_count": 256, "eval_duration_ns": 5725031000, "decode_tok_s": 44.716, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 31630000, "prefill_tok_s": 32342.713, "load_duration_ns": 194038666, "total_duration_ns": 5954571291, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.528, "decode_tok_s_wall_gross": 42.976, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.87, "median": 11.93, "max": 11.95, "mean": 11.92, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.92 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "power_w_all_cards": 11.92, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 138.3, "max_w": 141.18, "median_w": 140.03, "mean_w_whole_window": 118.85, "max_w_whole_window": 141.18, "limit_w": 150.0, "over_idle_w": 126.38, "samples": 22, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.83, "median": 139.78, "max": 141.18, "mean": 118.85, "n": 22 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 22 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 22 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 81.0, "n": 22 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.6, "n": 22 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1597.2, "n": 22 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13401.4, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 107.61, "median": 140.03, "max": 141.18, "mean": 138.3, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 45.0, "max": 46.0, "mean": 45.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1627.0, "median": 1642.0, "max": 1657.0, "mean": 1641.0, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3092.86, "power_mean_w_all_cards": 138.3, "energy_j_per_1k_tokens_all_cards": 3092.86, "watts_per_tok_s": 3.0929, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1627.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 138.3, "power_pct_of_limit": 92.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 92.2% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1627.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 138.3, "power_pct_of_limit": 92.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 92.2% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.7, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 198.83, "first_token_was_thinking": false, "wall_s": 5.924, "eval_count": 256, "eval_duration_ns": 5723589000, "decode_tok_s": 44.727, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27486000, "prefill_tok_s": 37218.948, "load_duration_ns": 166841147, "total_duration_ns": 5921849871, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.541, "decode_tok_s_wall_gross": 43.214, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.98, "median": 12.04, "max": 12.38, "mean": 12.13, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.13 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "power_w_all_cards": 12.13, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 134.7, "max_w": 141.31, "median_w": 140.84, "mean_w_whole_window": 119.73, "max_w_whole_window": 141.31, "limit_w": 150.0, "over_idle_w": 122.57, "samples": 22, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.76, "median": 140.8, "max": 141.31, "mean": 119.73, "n": 22 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 22 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 22 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.6, "n": 22 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1602.2, "n": 22 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13401.4, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 54.32, "median": 140.84, "max": 141.31, "mean": 134.7, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 44.0, "median": 45.0, "max": 46.0, "mean": 45.3, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1657.0, "mean": 1644.9, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 3011.59, "power_mean_w_all_cards": 134.7, "energy_j_per_1k_tokens_all_cards": 3011.59, "watts_per_tok_s": 3.0116, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.7, "power_pct_of_limit": 89.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.7, "power_pct_of_limit": 89.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.8, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 212.19, "first_token_was_thinking": false, "wall_s": 5.9298, "eval_count": 256, "eval_duration_ns": 5716297000, "decode_tok_s": 44.784, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27545000, "prefill_tok_s": 37139.227, "load_duration_ns": 178875799, "total_duration_ns": 5927125563, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.599, "decode_tok_s_wall_gross": 43.171, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.67, "median": 11.99, "max": 12.42, "mean": 12.03, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.03 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "power_w_all_cards": 12.03, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 135.14, "max_w": 141.23, "median_w": 140.85, "mean_w_whole_window": 120.12, "max_w_whole_window": 141.23, "limit_w": 150.0, "over_idle_w": 123.11, "samples": 22, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.69, "median": 140.59, "max": 141.23, "mean": 120.12, "n": 22 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 22 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 22 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.5, "n": 22 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1650.0, "max": 1740.0, "mean": 1610.2, "n": 22 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13401.4, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 58.36, "median": 140.85, "max": 141.23, "mean": 135.14, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15073.0, "median": 15073.0, "max": 15073.0, "mean": 15073.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 43.0, "median": 45.0, "max": 46.0, "mean": 45.1, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1647.4, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 3017.58, "power_mean_w_all_cards": 135.14, "energy_j_per_1k_tokens_all_cards": 3017.58, "watts_per_tok_s": 3.0176, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.14, "power_pct_of_limit": 90.1, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.1% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.14, "power_pct_of_limit": 90.1, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.1% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.7, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.716, "median": 44.727, "max": 44.784, "mean": 44.742, "n": 3 }, "ttft_ms": { "min": 198.83, "median": 212.19, "max": 230.08, "mean": 213.7, "n": 3 }, "power_mean_w": { "min": 134.7, "median": 135.14, "max": 138.3, "mean": 136.05, "n": 3 }, "power_mean_w_all_cards": { "min": 134.7, "median": 135.14, "max": 138.3, "mean": 136.05, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 3011.59, "median": 3017.58, "max": 3092.86, "mean": 3040.68, "n": 3 }, "energy_j_per_1k_tokens": { "min": 3011.59, "median": 3017.58, "max": 3092.86, "mean": 3040.68, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 678.7, "median": 678.8, "max": 679.7, "mean": 679.1, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 46.0, "max": 46.0, "mean": 46.0, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:27:39Z" }, { "num_ctx": 8192, "started_utc": "2026-09-22T00:27:39Z", "loads": true, "size_total": 15202789621, "size_vram": 15202789621, "context_length_loaded": 8192, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 588.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15073.0 }, "runner_peak_rss_gib": 0.86, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1821823, "rss_kb": 897432, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 28663 11130 4482 30978 34034\nSwap: 18431 18431 0", "kv_footprint_bytes": 599795958, "kv_footprint_gib": 0.56, "runs": [ { "ttft_ms": 181.74, "first_token_was_thinking": false, "wall_s": 5.897, "eval_count": 256, "eval_duration_ns": 5712580000, "decode_tok_s": 44.813, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 23033000, "prefill_tok_s": 44414.536, "load_duration_ns": 155357904, "total_duration_ns": 5893985506, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.617, "decode_tok_s_wall_gross": 43.412, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.83, "median": 22.85, "max": 22.86, "mean": 22.85, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.85 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "power_w_all_cards": 22.85, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 134.96, "max_w": 141.05, "median_w": 140.72, "mean_w_whole_window": 120.95, "max_w_whole_window": 141.05, "limit_w": 150.0, "over_idle_w": 112.11, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.83, "median": 140.66, "max": 141.05, "mean": 120.95, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 38.0, "median": 45.0, "max": 47.0, "mean": 44.9, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1590.0, "median": 1657.0, "max": 1755.0, "mean": 1651.9, "n": 21 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 21 }, "throttle_sw_power_cap_fraction": 0.8571, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 62.33, "median": 140.72, "max": 141.05, "mean": 134.96, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 46.0, "max": 47.0, "mean": 45.5, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1657.0, "max": 1657.0, "mean": 1653.0, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3011.6, "power_mean_w_all_cards": 134.96, "energy_j_per_1k_tokens_all_cards": 3011.6, "watts_per_tok_s": 3.0116, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1657.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.96, "power_pct_of_limit": 90.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1657.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.96, "power_pct_of_limit": 90.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 680.1, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 217.98, "first_token_was_thinking": false, "wall_s": 5.9396, "eval_count": 256, "eval_duration_ns": 5719872000, "decode_tok_s": 44.756, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27552000, "prefill_tok_s": 37129.791, "load_duration_ns": 187025864, "total_duration_ns": 5937566582, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.568, "decode_tok_s_wall_gross": 43.1, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 12.36, "median": 12.43, "max": 12.47, "mean": 12.42, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.42 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "power_w_all_cards": 12.42, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 134.51, "max_w": 140.93, "median_w": 140.44, "mean_w_whole_window": 118.99, "max_w_whole_window": 140.93, "limit_w": 150.0, "over_idle_w": 122.09, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.76, "median": 140.43, "max": 140.93, "mean": 118.99, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.5, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1755.0, "mean": 1608.6, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9524, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 54.78, "median": 140.44, "max": 140.93, "mean": 134.51, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 45.0, "max": 46.0, "mean": 45.2, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1635.0, "median": 1642.0, "max": 1657.0, "mean": 1646.7, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3005.39, "power_mean_w_all_cards": 134.51, "energy_j_per_1k_tokens_all_cards": 3005.39, "watts_per_tok_s": 3.0054, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.51, "power_pct_of_limit": 89.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.51, "power_pct_of_limit": 89.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.3, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 199.77, "first_token_was_thinking": false, "wall_s": 5.9313, "eval_count": 256, "eval_duration_ns": 5729463000, "decode_tok_s": 44.681, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27483000, "prefill_tok_s": 37223.011, "load_duration_ns": 168006577, "total_duration_ns": 5928572640, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.49, "decode_tok_s_wall_gross": 43.161, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.67, "median": 11.7, "max": 12.01, "mean": 11.79, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.79 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "power_w_all_cards": 11.79, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 16, "mean_w": 140.27, "max_w": 140.77, "median_w": 140.29, "mean_w_whole_window": 118.54, "max_w_whole_window": 140.77, "limit_w": 150.0, "over_idle_w": 128.48, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.89, "median": 140.13, "max": 140.77, "mean": 118.54, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 75.4, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.5, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1610.7, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 139.73, "median": 140.29, "max": 140.77, "mean": 140.27, "n": 16 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 16 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 16 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 15661.0, "median": 15661.0, "max": 15661.0, "mean": 15661.0, "n": 16 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 16 }, "temperature_c": { "min": 44.0, "median": 45.0, "max": 46.0, "mean": 45.3, "n": 16 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 16 }, "clocks_sm_mhz": { "min": 1635.0, "median": 1642.0, "max": 1657.0, "mean": 1642.1, "n": 16 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 16 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 16 }, "busy_samples": 16 } } }, "energy_j_per_1k_tokens": 3139.34, "power_mean_w_all_cards": 140.27, "energy_j_per_1k_tokens_all_cards": 3139.34, "watts_per_tok_s": 3.1394, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 140.27, "power_pct_of_limit": 93.5, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 93.5% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 140.27, "power_pct_of_limit": 93.5, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 93.5% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.1, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.681, "median": 44.756, "max": 44.813, "mean": 44.75, "n": 3 }, "ttft_ms": { "min": 181.74, "median": 199.77, "max": 217.98, "mean": 199.83, "n": 3 }, "power_mean_w": { "min": 134.51, "median": 134.96, "max": 140.27, "mean": 136.58, "n": 3 }, "power_mean_w_all_cards": { "min": 134.51, "median": 134.96, "max": 140.27, "mean": 136.58, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 3005.39, "median": 3011.6, "max": 3139.34, "mean": 3052.11, "n": 3 }, "energy_j_per_1k_tokens": { "min": 3005.39, "median": 3011.6, "max": 3139.34, "mean": 3052.11, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 678.1, "median": 679.3, "max": 680.1, "mean": 679.2, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 46.0, "max": 47.0, "mean": 46.3, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:29:06Z" }, { "num_ctx": 16384, "started_utc": "2026-09-22T00:29:06Z", "loads": true, "size_total": 15932598517, "size_vram": 15932598517, "context_length_loaded": 16384, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 696.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15661.0 }, "runner_peak_rss_gib": 0.87, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1827226, "rss_kb": 914080, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27121 12613 4498 31052 35577\nSwap: 18431 18431 0", "kv_footprint_bytes": 1329604854, "kv_footprint_gib": 1.24, "runs": [ { "ttft_ms": 215.47, "first_token_was_thinking": false, "wall_s": 5.9323, "eval_count": 256, "eval_duration_ns": 5714465000, "decode_tok_s": 44.799, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 22538000, "prefill_tok_s": 45390.008, "load_duration_ns": 189178368, "total_duration_ns": 5929471127, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.605, "decode_tok_s_wall_gross": 43.154, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.7, "median": 22.7, "max": 22.75, "mean": 22.72, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.72 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "power_w_all_cards": 22.72, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 20, "mean_w": 131.2, "max_w": 141.05, "median_w": 140.86, "mean_w_whole_window": 121.62, "max_w_whole_window": 141.05, "limit_w": 150.0, "over_idle_w": 108.48, "samples": 22, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.62, "median": 140.75, "max": 141.05, "mean": 121.62, "n": 22 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 22 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 22 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.8, "n": 22 }, "temperature_c": { "min": 38.0, "median": 46.0, "max": 47.0, "mean": 45.3, "n": 22 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 }, "clocks_sm_mhz": { "min": 1590.0, "median": 1650.0, "max": 1710.0, "mean": 1651.4, "n": 22 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 50.5, "median": 140.86, "max": 141.05, "mean": 131.2, "n": 20 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 20 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 20 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 20 }, "utilization_pct": { "min": 97.0, "median": 99.0, "max": 99.0, "mean": 98.8, "n": 20 }, "temperature_c": { "min": 44.0, "median": 46.0, "max": 47.0, "mean": 45.9, "n": 20 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 20 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1651.5, "n": 20 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 2928.66, "power_mean_w_all_cards": 131.2, "energy_j_per_1k_tokens_all_cards": 2928.66, "watts_per_tok_s": 2.9286, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 131.2, "power_pct_of_limit": 87.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 131.2, "power_pct_of_limit": 87.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.9, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 192.35, "first_token_was_thinking": false, "wall_s": 5.9186, "eval_count": 256, "eval_duration_ns": 5724817000, "decode_tok_s": 44.718, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27732000, "prefill_tok_s": 36888.793, "load_duration_ns": 160693930, "total_duration_ns": 5916272797, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.531, "decode_tok_s_wall_gross": 43.253, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.78, "median": 11.85, "max": 12.39, "mean": 12.01, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.01 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "power_w_all_cards": 12.01, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 129.35, "max_w": 141.13, "median_w": 140.61, "mean_w_whole_window": 118.79, "max_w_whole_window": 141.13, "limit_w": 150.0, "over_idle_w": 117.34, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 12.41, "median": 140.61, "max": 141.13, "mean": 118.79, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.6, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.7, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1601.3, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 39.61, "median": 140.61, "max": 141.13, "mean": 129.35, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 43.0, "median": 45.0, "max": 46.0, "mean": 45.3, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1657.0, "mean": 1644.4, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2892.6, "power_mean_w_all_cards": 129.35, "energy_j_per_1k_tokens_all_cards": 2892.6, "watts_per_tok_s": 2.8926, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.35, "power_pct_of_limit": 86.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.35, "power_pct_of_limit": 86.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.7, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 213.49, "first_token_was_thinking": false, "wall_s": 5.9297, "eval_count": 256, "eval_duration_ns": 5714288000, "decode_tok_s": 44.8, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20361000, "prefill_tok_s": 50243.112, "load_duration_ns": 189698209, "total_duration_ns": 5927560937, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.61, "decode_tok_s_wall_gross": 43.173, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.81, "median": 22.81, "max": 22.83, "mean": 22.82, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.82 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "power_w_all_cards": 22.82, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 139.76, "max_w": 141.37, "median_w": 141.09, "mean_w_whole_window": 121.91, "max_w_whole_window": 141.37, "limit_w": 150.0, "over_idle_w": 116.94, "samples": 22, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.76, "median": 140.91, "max": 141.37, "mean": 121.91, "n": 22 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 22 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 22 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 81.0, "n": 22 }, "temperature_c": { "min": 38.0, "median": 46.0, "max": 47.0, "mean": 45.5, "n": 22 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 }, "clocks_sm_mhz": { "min": 1590.0, "median": 1650.0, "max": 1725.0, "mean": 1648.7, "n": 22 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 118.04, "median": 141.09, "max": 141.37, "mean": 139.76, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 16357.0, "median": 16357.0, "max": 16357.0, "mean": 16357.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 45.0, "median": 46.0, "max": 47.0, "mean": 46.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1651.4, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3119.64, "power_mean_w_all_cards": 139.76, "energy_j_per_1k_tokens_all_cards": 3119.64, "watts_per_tok_s": 3.1196, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 45.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 139.76, "power_pct_of_limit": 93.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 93.2% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 45.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 139.76, "power_pct_of_limit": 93.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (draw held at 93.2% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.9, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.718, "median": 44.799, "max": 44.8, "mean": 44.772, "n": 3 }, "ttft_ms": { "min": 192.35, "median": 213.49, "max": 215.47, "mean": 207.1, "n": 3 }, "power_mean_w": { "min": 129.35, "median": 131.2, "max": 139.76, "mean": 133.44, "n": 3 }, "power_mean_w_all_cards": { "min": 129.35, "median": 131.2, "max": 139.76, "mean": 133.44, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 2892.6, "median": 2928.66, "max": 3119.64, "mean": 2980.3, "n": 3 }, "energy_j_per_1k_tokens": { "min": 2892.6, "median": 2928.66, "max": 3119.64, "mean": 2980.3, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 678.7, "median": 679.9, "max": 679.9, "mean": 679.5, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 47.0, "max": 47.0, "mean": 46.7, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:30:17Z" }, { "num_ctx": 32768, "started_utc": "2026-09-22T00:30:17Z", "loads": true, "size_total": 17127964671, "size_vram": 17127964671, "context_length_loaded": 32768, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 1124.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 16357.0 }, "runner_peak_rss_gib": 0.85, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1836175, "rss_kb": 892572, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27186 12523 4478 31058 35512\nSwap: 18431 18431 0", "kv_footprint_bytes": 2524971008, "kv_footprint_gib": 2.35, "runs": [ { "ttft_ms": 202.6, "first_token_was_thinking": false, "wall_s": 5.9154, "eval_count": 256, "eval_duration_ns": 5710752000, "decode_tok_s": 44.828, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 23278000, "prefill_tok_s": 43947.074, "load_duration_ns": 175724043, "total_duration_ns": 5912786425, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.637, "decode_tok_s_wall_gross": 43.277, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.81, "median": 22.83, "max": 22.86, "mean": 22.83, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.83 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "power_w_all_cards": 22.83, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 130.88, "max_w": 141.51, "median_w": 141.29, "mean_w_whole_window": 121.15, "max_w_whole_window": 141.51, "limit_w": 150.0, "over_idle_w": 108.05, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.81, "median": 141.26, "max": 141.51, "mean": 121.15, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.6, "n": 21 }, "temperature_c": { "min": 39.0, "median": 47.0, "max": 48.0, "mean": 45.9, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1590.0, "median": 1650.0, "max": 1657.0, "mean": 1646.2, "n": 21 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 47.9, "median": 141.29, "max": 141.51, "mean": 130.88, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 44.0, "median": 47.0, "max": 48.0, "mean": 46.5, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1652.1, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2919.62, "power_mean_w_all_cards": 130.88, "energy_j_per_1k_tokens_all_cards": 2919.62, "watts_per_tok_s": 2.9196, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 48.0, "temp_min_c": 44.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 130.88, "power_pct_of_limit": 87.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 48.0, "temp_min_c": 44.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 130.88, "power_pct_of_limit": 87.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 680.4, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 171.39, "first_token_was_thinking": false, "wall_s": 5.8882, "eval_count": 256, "eval_duration_ns": 5714915000, "decode_tok_s": 44.795, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27948000, "prefill_tok_s": 36603.693, "load_duration_ns": 140137532, "total_duration_ns": 5886177252, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.605, "decode_tok_s_wall_gross": 43.477, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.72, "median": 11.72, "max": 12.32, "mean": 11.92, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.92 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "power_w_all_cards": 11.92, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 135.59, "max_w": 141.51, "median_w": 141.14, "mean_w_whole_window": 120.13, "max_w_whole_window": 141.51, "limit_w": 150.0, "over_idle_w": 123.67, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 12.1, "median": 141.08, "max": 141.51, "mean": 120.13, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 38.0, "median": 46.0, "max": 47.0, "mean": 45.3, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1650.0, "max": 1770.0, "mean": 1611.2, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.8571, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 59.08, "median": 141.14, "max": 141.51, "mean": 135.59, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 46.0, "max": 47.0, "mean": 46.0, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1648.9, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3026.9, "power_mean_w_all_cards": 135.59, "energy_j_per_1k_tokens_all_cards": 3026.9, "watts_per_tok_s": 3.0269, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.59, "power_pct_of_limit": 90.4, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.4% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.59, "power_pct_of_limit": 90.4, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.4% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.9, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 193.6, "first_token_was_thinking": false, "wall_s": 5.9144, "eval_count": 256, "eval_duration_ns": 5719304000, "decode_tok_s": 44.761, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27879000, "prefill_tok_s": 36694.286, "load_duration_ns": 161677569, "total_duration_ns": 5912136884, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.575, "decode_tok_s_wall_gross": 43.285, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.59, "median": 12.03, "max": 12.33, "mean": 11.98, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.98 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "power_w_all_cards": 11.98, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 135.2, "max_w": 141.49, "median_w": 140.91, "mean_w_whole_window": 119.55, "max_w_whole_window": 141.49, "limit_w": 150.0, "over_idle_w": 123.22, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.65, "median": 140.69, "max": 141.49, "mean": 119.55, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 37.0, "median": 46.0, "max": 47.0, "mean": 45.0, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1616.0, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.8571, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 56.65, "median": 140.91, "max": 141.49, "mean": 135.2, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 17481.0, "median": 17481.0, "max": 17481.0, "mean": 17481.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 46.0, "max": 47.0, "mean": 45.7, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1657.0, "mean": 1648.8, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3020.51, "power_mean_w_all_cards": 135.2, "energy_j_per_1k_tokens_all_cards": 3020.51, "watts_per_tok_s": 3.0205, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.2, "power_pct_of_limit": 90.1, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.1% of the 150 W cap)" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 135.2, "power_pct_of_limit": 90.1, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode (draw held at 90.1% of the 150 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.4, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.761, "median": 44.795, "max": 44.828, "mean": 44.795, "n": 3 }, "ttft_ms": { "min": 171.39, "median": 193.6, "max": 202.6, "mean": 189.2, "n": 3 }, "power_mean_w": { "min": 130.88, "median": 135.2, "max": 135.59, "mean": 133.89, "n": 3 }, "power_mean_w_all_cards": { "min": 130.88, "median": 135.2, "max": 135.59, "mean": 133.89, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 2919.62, "median": 3020.51, "max": 3026.9, "mean": 2989.01, "n": 3 }, "energy_j_per_1k_tokens": { "min": 2919.62, "median": 3020.51, "max": 3026.9, "mean": 2989.01, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 679.4, "median": 679.9, "max": 680.4, "mean": 679.9, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 47.0, "median": 47.0, "max": 48.0, "mean": 47.3, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:31:39Z" }, { "num_ctx": 49152, "started_utc": "2026-09-22T00:31:39Z", "loads": true, "size_total": 18579193855, "size_vram": 18579193855, "context_length_loaded": 49152, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 1384.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 17481.0 }, "runner_peak_rss_gib": 0.87, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1849543, "rss_kb": 909640, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27337 10796 4494 32650 35360\nSwap: 18431 18431 0", "kv_footprint_bytes": 3976200192, "kv_footprint_gib": 3.7, "runs": [ { "ttft_ms": 199.37, "first_token_was_thinking": false, "wall_s": 5.9144, "eval_count": 256, "eval_duration_ns": 5713268000, "decode_tok_s": 44.808, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 23134000, "prefill_tok_s": 44220.628, "load_duration_ns": 172159547, "total_duration_ns": 5911831550, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.619, "decode_tok_s_wall_gross": 43.284, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.67, "median": 22.69, "max": 22.76, "mean": 22.71, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.71 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "power_w_all_cards": 22.71, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 130.81, "max_w": 141.36, "median_w": 140.94, "mean_w_whole_window": 120.75, "max_w_whole_window": 141.36, "limit_w": 150.0, "over_idle_w": 108.1, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.72, "median": 140.93, "max": 141.36, "mean": 120.75, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.6, "n": 21 }, "temperature_c": { "min": 38.0, "median": 46.0, "max": 47.0, "mean": 45.4, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1590.0, "median": 1650.0, "max": 1755.0, "mean": 1649.0, "n": 21 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 49.12, "median": 140.94, "max": 141.36, "mean": 130.81, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 44.0, "median": 46.0, "max": 47.0, "mean": 46.0, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1650.0, "max": 1755.0, "mean": 1655.3, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2919.35, "power_mean_w_all_cards": 130.81, "energy_j_per_1k_tokens_all_cards": 2919.35, "watts_per_tok_s": 2.9193, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1755.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": true, "power_limit_w": 150.0, "power_mean_w": 130.81, "power_pct_of_limit": 87.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 87.2% of the cap and the card at 47 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 47.0, "temp_min_c": 44.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1755.0, "clock_median_mhz": 1650.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": true, "power_limit_w": 150.0, "power_mean_w": 130.81, "power_pct_of_limit": 87.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 87.2% of the cap and the card at 47 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 680.1, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 202.19, "first_token_was_thinking": false, "wall_s": 5.9227, "eval_count": 256, "eval_duration_ns": 5718733000, "decode_tok_s": 44.765, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28066000, "prefill_tok_s": 36449.797, "load_duration_ns": 170177068, "total_duration_ns": 5920195115, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.577, "decode_tok_s_wall_gross": 43.224, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.68, "median": 11.76, "max": 11.84, "mean": 11.76, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.76 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "power_w_all_cards": 11.76, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 129.37, "max_w": 141.27, "median_w": 140.6, "mean_w_whole_window": 118.66, "max_w_whole_window": 141.27, "limit_w": 150.0, "over_idle_w": 117.61, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.75, "median": 140.6, "max": 141.27, "mean": 118.66, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.6, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 47.0, "mean": 44.8, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1717.0, "mean": 1608.6, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 42.99, "median": 140.6, "max": 141.27, "mean": 129.37, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 43.0, "median": 46.0, "max": 47.0, "mean": 45.4, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1635.0, "median": 1642.0, "max": 1717.0, "mean": 1650.1, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2889.97, "power_mean_w_all_cards": 129.37, "energy_j_per_1k_tokens_all_cards": 2889.97, "watts_per_tok_s": 2.89, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 47.0, "temp_min_c": 43.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1717.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.37, "power_pct_of_limit": 86.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 47.0, "temp_min_c": 43.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1635.0, "clock_max_mhz": 1717.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.37, "power_pct_of_limit": 86.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.4, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 191.25, "first_token_was_thinking": false, "wall_s": 5.9141, "eval_count": 256, "eval_duration_ns": 5720483000, "decode_tok_s": 44.751, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27667000, "prefill_tok_s": 36975.458, "load_duration_ns": 159616741, "total_duration_ns": 5910996234, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.558, "decode_tok_s_wall_gross": 43.286, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.72, "median": 11.75, "max": 12.31, "mean": 11.93, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.93 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "power_w_all_cards": 11.93, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 134.81, "max_w": 141.23, "median_w": 140.74, "mean_w_whole_window": 119.23, "max_w_whole_window": 141.23, "limit_w": 150.0, "over_idle_w": 122.88, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.68, "median": 140.67, "max": 141.23, "mean": 119.23, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.5, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1616.0, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.8571, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 58.61, "median": 140.74, "max": 141.23, "mean": 134.81, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 18865.0, "median": 18865.0, "max": 18865.0, "mean": 18865.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 44.0, "median": 45.0, "max": 46.0, "mean": 45.2, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1657.0, "mean": 1646.7, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 3012.42, "power_mean_w_all_cards": 134.81, "energy_j_per_1k_tokens_all_cards": 3012.42, "watts_per_tok_s": 3.0124, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.81, "power_pct_of_limit": 89.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 44.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 134.81, "power_pct_of_limit": 89.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 679.2, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.751, "median": 44.765, "max": 44.808, "mean": 44.775, "n": 3 }, "ttft_ms": { "min": 191.25, "median": 199.37, "max": 202.19, "mean": 197.6, "n": 3 }, "power_mean_w": { "min": 129.37, "median": 130.81, "max": 134.81, "mean": 131.66, "n": 3 }, "power_mean_w_all_cards": { "min": 129.37, "median": 130.81, "max": 134.81, "mean": 131.66, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 2889.97, "median": 2919.35, "max": 3012.42, "mean": 2940.58, "n": 3 }, "energy_j_per_1k_tokens": { "min": 2889.97, "median": 2919.35, "max": 3012.42, "mean": 2940.58, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 679.2, "median": 679.4, "max": 680.1, "mean": 679.6, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 47.0, "max": 47.0, "mean": 46.7, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:33:01Z" }, { "num_ctx": 65536, "started_utc": "2026-09-22T00:33:01Z", "loads": true, "size_total": 20089143295, "size_vram": 20089143295, "context_length_loaded": 65536, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 1440.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 18865.0 }, "runner_peak_rss_gib": 0.88, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 1853026, "rss_kb": 926144, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27371 10634 4510 32794 35327\nSwap: 18431 18431 0", "kv_footprint_bytes": 5486149632, "kv_footprint_gib": 5.11, "runs": [ { "ttft_ms": 206.28, "first_token_was_thinking": false, "wall_s": 5.9348, "eval_count": 256, "eval_duration_ns": 5726824000, "decode_tok_s": 44.702, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 31687000, "prefill_tok_s": 32284.533, "load_duration_ns": 170860022, "total_duration_ns": 5932461585, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.514, "decode_tok_s_wall_gross": 43.135, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.69, "median": 12.33, "max": 12.73, "mean": 12.25, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.25 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "power_w_all_cards": 12.25, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 133.71, "max_w": 140.6, "median_w": 140.31, "mean_w_whole_window": 118.22, "max_w_whole_window": 140.6, "limit_w": 150.0, "over_idle_w": 121.46, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.65, "median": 140.18, "max": 140.6, "mean": 118.22, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.9, "n": 21 }, "temperature_c": { "min": 37.0, "median": 45.0, "max": 46.0, "mean": 44.3, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1695.0, "mean": 1604.6, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9524, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 52.77, "median": 140.31, "max": 140.6, "mean": 133.71, "n": 18 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 18 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 18 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 43.0, "median": 45.0, "max": 46.0, "mean": 44.9, "n": 18 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 18 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1657.0, "mean": 1645.4, "n": 18 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 2991.15, "power_mean_w_all_cards": 133.71, "energy_j_per_1k_tokens_all_cards": 2991.15, "watts_per_tok_s": 2.9911, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 133.71, "power_pct_of_limit": 89.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 46.0, "temp_min_c": 43.0, "temp_rise_c": 3.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 133.71, "power_pct_of_limit": 89.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.5, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 191.65, "first_token_was_thinking": false, "wall_s": 5.9234, "eval_count": 256, "eval_duration_ns": 5729691000, "decode_tok_s": 44.68, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27972000, "prefill_tok_s": 36572.287, "load_duration_ns": 159612319, "total_duration_ns": 5920706669, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.489, "decode_tok_s_wall_gross": 43.219, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.72, "median": 11.73, "max": 11.81, "mean": 11.75, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 11.75 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "power_w_all_cards": 11.75, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 129.41, "max_w": 140.66, "median_w": 140.3, "mean_w_whole_window": 118.53, "max_w_whole_window": 140.66, "limit_w": 150.0, "over_idle_w": 117.66, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 11.72, "median": 140.24, "max": 140.66, "mean": 118.53, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.4, "n": 21 }, "temperature_c": { "min": 36.0, "median": 45.0, "max": 46.0, "mean": 44.0, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1702.0, "mean": 1604.9, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 41.11, "median": 140.3, "max": 140.66, "mean": 129.41, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 19 }, "utilization_pct": { "min": 98.0, "median": 99.0, "max": 99.0, "mean": 98.8, "n": 19 }, "temperature_c": { "min": 42.0, "median": 45.0, "max": 46.0, "mean": 44.7, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1702.0, "mean": 1645.2, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2896.4, "power_mean_w_all_cards": 129.41, "energy_j_per_1k_tokens_all_cards": 2896.4, "watts_per_tok_s": 2.8964, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 42.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1702.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.41, "power_pct_of_limit": 86.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 42.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1702.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.41, "power_pct_of_limit": 86.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.1, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 184.56, "first_token_was_thinking": false, "wall_s": 5.9105, "eval_count": 256, "eval_duration_ns": 5723741000, "decode_tok_s": 44.726, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27958000, "prefill_tok_s": 36590.6, "load_duration_ns": 152391836, "total_duration_ns": 5907832922, "streamed_chunks": 256, "response_chars": 1022, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 44.534, "decode_tok_s_wall_gross": 43.313, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 11.89, "median": 11.91, "max": 12.93, "mean": 12.24, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 12.24 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "power_w_all_cards": 12.24, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 129.9, "max_w": 140.71, "median_w": 140.22, "mean_w_whole_window": 118.9, "max_w_whole_window": 140.71, "limit_w": 150.0, "over_idle_w": 117.66, "samples": 21, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 12.29, "median": 140.2, "max": 140.71, "mean": 118.9, "n": 21 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 21 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 21 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.6, "n": 21 }, "temperature_c": { "min": 36.0, "median": 45.0, "max": 46.0, "mean": 44.0, "n": 21 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 }, "clocks_sm_mhz": { "min": 1192.0, "median": 1642.0, "max": 1657.0, "mean": 1616.0, "n": 21 }, "clocks_mem_mhz": { "min": 810.0, "median": 14001.0, "max": 14001.0, "mean": 13372.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 41.75, "median": 140.22, "max": 140.71, "mean": 129.9, "n": 19 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "shipped", "power_posture": "shipped", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 19 }, "power_posture_note": "shipped -- nvidia-powerd running and no cap set: the board floats between its power.default_limit and its power.max_limit against the CPU's draw, DURING the scored run. `power_limit_enforced_stats_w` is that limit reduced over this run's own 2 Hz trace and is the only honest cap cell; `power_limit_enforced_w` is the FIRST sample and is kept only so the two can be compared. A single wattage in a cap column for this posture is a fabrication. AND this box's boot-time clock lock (ai-perf.service, -lgc 1200,2550) is IN FORCE, which is this posture's declared condition rather than a confound: it is how the machine boots. The verdict is read FROM THE CARD and lives in the pass's clock-lock receipt. Rows measured here are comparable with this machine's other two postures and with nothing else in the estate -- PREREG.md Amendment 4.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 19 samples of this run's 2 Hz trace. Under shipped that is a FINDING, not a range.", "memory_used_mib": { "min": 20305.0, "median": 20305.0, "max": 20305.0, "mean": 20305.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 42.0, "median": 45.0, "max": 46.0, "mean": 44.6, "n": 19 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 19 }, "clocks_sm_mhz": { "min": 1642.0, "median": 1642.0, "max": 1657.0, "mean": 1644.8, "n": 19 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 2904.35, "power_mean_w_all_cards": 129.9, "energy_j_per_1k_tokens_all_cards": 2904.35, "watts_per_tok_s": 2.9044, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 42.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.9, "power_pct_of_limit": 86.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 46.0, "temp_min_c": 42.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 8.0, "pcie_width_max": 8.0, "pcie_width_median": 8.0, "fan_mean_pct": null, "fan_max_pct": null, "clock_floor_mhz": 1642.0, "clock_max_mhz": 1657.0, "clock_median_mhz": 1642.0, "mem_clock_floor_mhz": 14001.0, "mem_clock_median_mhz": 14001.0, "mem_clock_max_mhz": 14001.0, "mem_clock_held": true, "memory_bus_width_bits": 256, "memory_technology": "GDDR7", "memory_bits_per_clock": null, "peak_mem_bandwidth_gbs_at_median_clock": null, "peak_mem_bandwidth_gbs_at_floor_clock": null, "peak_mem_bandwidth_gbs_at_max_clock": null, "peak_mem_bandwidth_withheld_because": "this harness's peak-bandwidth formula assumes two bits per memory clock per pin, which is exact for GDDR6 and GDDR6X and is checked against three published bandwidths. It has no receipt for the bits-per-clock of this board's memory, so the DERIVED bytes-per-second ceiling is withheld rather than computed from an unverified constant. The memory clock itself is a reading and is published unchanged.", "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate. The bits-per-clock is registered per board and is NOT defaulted.", "clock_dropped": false, "power_limit_w": 150.0, "power_mean_w": 129.9, "power_pct_of_limit": 86.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": 678.8, "fraction_of_peak_pct": null, "basis": "decode tok/s x the model's own store size in bytes: a dense transformer reads every weight once per decoded token. A FLOOR on the traffic -- the KV cache is read on top of it -- and it uses the store size, which is the weights plus a few kilobytes of template and licence.", "refused_reason": null }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 44.68, "median": 44.702, "max": 44.726, "mean": 44.703, "n": 3 }, "ttft_ms": { "min": 184.56, "median": 191.65, "max": 206.28, "mean": 194.16, "n": 3 }, "power_mean_w": { "min": 129.41, "median": 129.9, "max": 133.71, "mean": 131.01, "n": 3 }, "power_mean_w_all_cards": { "min": 129.41, "median": 129.9, "max": 133.71, "mean": 131.01, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 2896.4, "median": 2904.35, "max": 2991.15, "mean": 2930.63, "n": 3 }, "energy_j_per_1k_tokens": { "min": 2896.4, "median": 2904.35, "max": 2991.15, "mean": 2930.63, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": null, "decode_gbs": { "min": 678.1, "median": 678.5, "max": 678.8, "mean": 678.5, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 46.0, "max": 46.0, "mean": 46.0, "n": 3 }, "fan_max_pct": null }, "finished_utc": "2026-09-22T00:34:33Z" }, { "num_ctx": 98304, "started_utc": "2026-09-22T00:34:33Z", "loads": true, "size_total": 24696898560, "size_vram": 22763557355, "context_length_loaded": 98304, "fits_fully_in_vram": false, "verdict": "does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM)", "offload_split": "92.2% VRAM / 7.8% RAM", "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 2550.0 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22855.0 }, "memory_used_mib_per_card_before": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20305.0 }, "runner_peak_rss_gib": 14.17, "runner_rss_within_allowance": false, "runner_procs": [ { "pid": 1856825, "rss_kb": 14862196, "threads": 31 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 62698 27483 10423 4542 32925 35214\nSwap: 18431 18431 0", "kv_footprint_bytes": 10093904897, "kv_footprint_gib": 9.4, "finished_utc": "2026-09-22T00:34:38Z" } ], "errors": [], "context_length_max": 131072, "quantization": "Q4_K_M", "model_store_size_bytes": 15177384862, "model_store_size_gib": 14.14, "model_density": "dense", "ladder_stopped_at": { "num_ctx": 98304, "why": "the first rung that did not fit whole on the card", "verdict": "does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM)", "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22855.0 } }, "ups_after_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-22T00:34:39Z", "available": false, "error": "no upsc on this box" }, "largest_context_that_fits": 65536, "levels_requested": [ 4096, 8192, 16384, 32768, 49152, 65536, 98304, 131072 ], "levels_not_reached": [ 131072 ], "headline": "the biggest context this arm holds for mistral-small3.2:24b is 65,536 tokens", "finished_utc": "2026-09-22T00:34:39Z" }