{ "model_id": "m5", "model_tag": "gemma4:26b", "box_class": "gpu-5090-laptop-24g", "arm": "one", "arm_label": "the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it", "arm_cards": [ "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff" ], "target_uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "live_seat_arm": false, "base_url": "http://127.0.0.1:11470", "started_utc": "2026-09-21T21:10:49Z", "nproc": 24, "num_predict": 256, "num_predict_is_series_default": true, "num_ctx_option": 131072, "prompt_path": "/workshop/bench-laptop-5090-2026-09-21/harness-llm/inputs/frozen-p512.txt", "prompt_sha256": "90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c", "prompt_bytes": 2101, "scored_runs_requested": 3, "warmup_discarded_runs": 1, "wall_decode_note": "decode_tok_s is ollama's own eval_count/eval_duration. decode_tok_s_wall is (eval_count - 1) / (wall_s - ttft_s): the tokens AFTER the first, over the time after the first, measured by this client's clock -- it carries the HTTP hop and the stream decode that ollama's counter does not. decode_tok_s_wall_gross is eval_count / wall_s and therefore includes prefill and any load; it is the slowest and least comparable of the three, printed so nobody has to derive it.", "ttft_note": "Every scored run sends the SAME frozen prompt, so from the second request the KV prefix is cached and derived.ttft_ms is a CACHED-PREFIX first-token time, not a fresh-prompt one. The CPU bench's LEDGER.md, 'THE HEADLINE TRAP'.", "loadavg_before_arm": "2.21 1.96 1.97 2/2634 1338285", "free_m_before": "total used free shared buff/cache available\nMem: 62698 27607 15220 4737 27816 35091\nSwap: 18431 18431 0", "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.", "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.", "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": "2026-09-21T21:10:49Z", "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." }, "instance_env": { "kind": "instance-env-receipt", "box_class": "gpu-5090-laptop-24g", "target_uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "written_utc": "2026-09-21T21:10:49Z", "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": "f16", "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, 1590 MHz, 14001 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;" }, "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.", "ups_before_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-21T21:10:49Z", "available": false, "error": "no upsc on this box" }, "runs": [ { "ttft_ms": 451.55, "first_token_was_thinking": false, "wall_s": 1.9951, "eval_count": 256, "eval_duration_ns": 1540540000, "decode_tok_s": 166.175, "prompt_eval_count": 536, "prompt_eval_duration_ns": 108491000, "prefill_tok_s": 4940.502, "load_duration_ns": 337668898, "total_duration_ns": 1991470635, "streamed_chunks": 256, "response_chars": 1064, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 165.208, "decode_tok_s_wall_gross": 128.317, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.47, "median": 22.5, "max": 22.51, "mean": 22.49, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.49 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "power_w_all_cards": 22.49, "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": 5, "mean_w": 114.49, "max_w": 147.7, "median_w": 141.76, "mean_w_whole_window": 80.38, "max_w_whole_window": 147.7, "limit_w": 150.0, "over_idle_w": 92.0, "samples": 8, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.47, "median": 68.28, "max": 147.7, "mean": 80.38, "n": 8 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 8 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 8 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 8 }, "utilization_pct": { "min": 0.0, "median": 93.5, "max": 94.0, "mean": 58.6, "n": 8 }, "temperature_c": { "min": 35.0, "median": 42.5, "max": 44.0, "mean": 40.5, "n": 8 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 }, "clocks_sm_mhz": { "min": 1590.0, "median": 2174.5, "max": 2265.0, "mean": 1990.1, "n": 8 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 8 }, "throttle_sw_power_cap_fraction": 0.625, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 8, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 45.53, "median": 141.76, "max": 147.7, "mean": 114.49, "n": 5 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 5 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 5 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 5 }, "utilization_pct": { "min": 93.0, "median": 94.0, "max": 94.0, "mean": 93.8, "n": 5 }, "temperature_c": { "min": 42.0, "median": 43.0, "max": 44.0, "mean": 43.0, "n": 5 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 5 }, "clocks_sm_mhz": { "min": 2122.0, "median": 2250.0, "max": 2265.0, "mean": 2224.2, "n": 5 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 5 }, "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": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 688.97, "power_mean_w_all_cards": 114.49, "energy_j_per_1k_tokens_all_cards": 688.97, "watts_per_tok_s": 0.689, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 44.0, "temp_min_c": 42.0, "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": 2122.0, "clock_max_mhz": 2265.0, "clock_median_mhz": 2250.0, "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": 114.49, "power_pct_of_limit": 76.3, "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": false, "verdict": "NEITHER cap: the SM clock varied with draw at 76.3% of the cap and the card at 44 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": 5, "temp_max_c": 44.0, "temp_min_c": 42.0, "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": 2122.0, "clock_max_mhz": 2265.0, "clock_median_mhz": 2250.0, "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": 114.49, "power_pct_of_limit": 76.3, "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": false, "verdict": "NEITHER cap: the SM clock varied with draw at 76.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 } }, "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": 436.29, "first_token_was_thinking": false, "wall_s": 1.9862, "eval_count": 256, "eval_duration_ns": 1545369000, "decode_tok_s": 165.656, "prompt_eval_count": 536, "prompt_eval_duration_ns": 110812000, "prefill_tok_s": 4837.021, "load_duration_ns": 320139305, "total_duration_ns": 1981122059, "streamed_chunks": 256, "response_chars": 1064, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 164.53, "decode_tok_s_wall_gross": 128.892, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.45, "median": 22.5, "max": 22.5, "mean": 22.48, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.48 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "power_w_all_cards": 22.48, "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": 5, "mean_w": 116.01, "max_w": 146.73, "median_w": 142.23, "mean_w_whole_window": 82.39, "max_w_whole_window": 146.73, "limit_w": 150.0, "over_idle_w": 93.53, "samples": 8, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.46, "median": 72.52, "max": 146.73, "mean": 82.39, "n": 8 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 8 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 8 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 8 }, "utilization_pct": { "min": 0.0, "median": 94.0, "max": 94.0, "mean": 58.8, "n": 8 }, "temperature_c": { "min": 35.0, "median": 42.5, "max": 44.0, "mean": 40.6, "n": 8 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 }, "clocks_sm_mhz": { "min": 1590.0, "median": 2175.0, "max": 2287.0, "mean": 1994.9, "n": 8 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 8 }, "throttle_sw_power_cap_fraction": 0.625, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 8, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 47.15, "median": 142.23, "max": 146.73, "mean": 116.01, "n": 5 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 5 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 5 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 5 }, "utilization_pct": { "min": 94.0, "median": 94.0, "max": 94.0, "mean": 94.0, "n": 5 }, "temperature_c": { "min": 42.0, "median": 43.0, "max": 44.0, "mean": 43.2, "n": 5 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 5 }, "clocks_sm_mhz": { "min": 2115.0, "median": 2257.0, "max": 2287.0, "mean": 2231.8, "n": 5 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 5 }, "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": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 700.31, "power_mean_w_all_cards": 116.01, "energy_j_per_1k_tokens_all_cards": 700.31, "watts_per_tok_s": 0.7003, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 44.0, "temp_min_c": 42.0, "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": 2115.0, "clock_max_mhz": 2287.0, "clock_median_mhz": 2257.0, "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": 116.01, "power_pct_of_limit": 77.3, "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": false, "verdict": "NEITHER cap: the SM clock varied with draw at 77.3% of the cap and the card at 44 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": 5, "temp_max_c": 44.0, "temp_min_c": 42.0, "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": 2115.0, "clock_max_mhz": 2287.0, "clock_median_mhz": 2257.0, "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": 116.01, "power_pct_of_limit": 77.3, "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": false, "verdict": "NEITHER cap: the SM clock varied with draw at 77.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 } }, "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": 439.78, "first_token_was_thinking": false, "wall_s": 1.9889, "eval_count": 256, "eval_duration_ns": 1546404000, "decode_tok_s": 165.545, "prompt_eval_count": 536, "prompt_eval_duration_ns": 116526000, "prefill_tok_s": 4599.832, "load_duration_ns": 318106357, "total_duration_ns": 1985201125, "streamed_chunks": 256, "response_chars": 1064, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 164.615, "decode_tok_s_wall_gross": 128.718, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.43, "median": 22.48, "max": 22.49, "mean": 22.47, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 6 }, "power_w_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 22.47 }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "power_w_all_cards": 22.47, "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": 3, "mean_w": 144.06, "max_w": 147.54, "median_w": 146.5, "mean_w_whole_window": 81.31, "max_w_whole_window": 147.54, "limit_w": 150.0, "over_idle_w": 121.59, "samples": 8, "per_gpu": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 22.41, "median": 73.23, "max": 147.54, "mean": 81.31, "n": 8 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 8 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 8 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 8 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 94.0, "mean": 35.2, "n": 8 }, "temperature_c": { "min": 36.0, "median": 42.5, "max": 44.0, "mean": 40.8, "n": 8 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 }, "clocks_sm_mhz": { "min": 1590.0, "median": 2178.5, "max": 2287.0, "mean": 1997.6, "n": 8 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 8 }, "throttle_sw_power_cap_fraction": 0.625, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 8, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "power_w": { "min": 138.15, "median": 146.5, "max": 147.54, "mean": 144.06, "n": 3 }, "power_limit_w": 150.0, "power_limit_field": "enforced.power.limit", "power_limit_enforced_w": 150.0, "cap": "dynamic-boost", "power_posture": "dynamic-boost", "power_limit_enforced_stats_w": { "min": 150.0, "median": 150.0, "max": 150.0, "mean": 150.0, "n": 3 }, "power_posture_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, 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.", "cap_cell": "150 W, flat", "cap_cell_note": "the enforced limit read 150 W on every one of 3 samples of this run's 2 Hz trace. Under dynamic-boost that is a FINDING, not a range.", "memory_used_mib": { "min": 20711.0, "median": 20711.0, "max": 20711.0, "mean": 20711.0, "n": 3 }, "utilization_pct": { "min": 94.0, "median": 94.0, "max": 94.0, "mean": 94.0, "n": 3 }, "temperature_c": { "min": 43.0, "median": 43.0, "max": 44.0, "mean": 43.3, "n": 3 }, "memory_temperature_c": null, "fan_pct": null, "pcie_width_current": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 3 }, "clocks_sm_mhz": { "min": 2107.0, "median": 2250.0, "max": 2265.0, "mean": 2207.3, "n": 3 }, "clocks_mem_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "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": 3 }, "busy_samples": 3 } } }, "energy_j_per_1k_tokens": 870.21, "power_mean_w_all_cards": 144.06, "energy_j_per_1k_tokens_all_cards": 870.21, "watts_per_tok_s": 0.8702, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 3, "temp_max_c": 44.0, "temp_min_c": 43.0, "temp_rise_c": 1.0, "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": 2107.0, "clock_max_mhz": 2265.0, "clock_median_mhz": 2250.0, "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": 144.06, "power_pct_of_limit": 96.0, "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": "the POWER CAP doing its job: the SM clock fell while draw sat at 96.0% of the 150 W cap, at 44 C -- well below any throttling temperature" }, "thermal_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 3, "temp_max_c": 44.0, "temp_min_c": 43.0, "temp_rise_c": 1.0, "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": 2107.0, "clock_max_mhz": 2265.0, "clock_median_mhz": 2250.0, "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": 144.06, "power_pct_of_limit": 96.0, "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": "the POWER CAP doing its job: the SM clock fell while draw sat at 96.0% of the 150 W cap, at 44 C -- well below any throttling temperature" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": null, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "pcie_width_under_load": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "min": 8.0, "median": 8.0, "max": 8.0, "mean": 8.0, "n": 8 } }, "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 } ], "errors": [], "nvidia_before": "0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, 15, 23970, 24463, 22.94, [N/A], 150.00, 35, N/A, [N/A], 1590, 14001, 0, 8, Not Active, Not Active, Not Active, Not Active", "smi_before": [ { "index": "0", "uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "memory.used": 15.0, "memory.free": 23970.0, "memory.total": 24463.0, "power.draw": 22.94, "power.limit": null, "enforced.power.limit": 150.0, "temperature.gpu": 35.0, "temperature.memory": null, "fan.speed": null, "clocks.sm": 1590.0, "clocks.mem": 14001.0, "utilization.gpu": 0.0, "pcie.link.width.current": 8.0, "clocks_throttle_reasons.sw_power_cap": "Not Active", "clocks_throttle_reasons.hw_slowdown": "Not Active", "clocks_throttle_reasons.sw_thermal_slowdown": "Not Active", "clocks_throttle_reasons.hw_thermal_slowdown": "Not Active" } ], "power_limit_w_at_start": null, "ollama_version": "0.32.13", "model_store_size_bytes": 17987581215, "model_store_size_gib": 16.75, "model_density": "moe", "vram_before_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 15.0 }, "quantization": "Q4_K_M", "param_size": "25.8B", "family": "gemma4", "context_length_max": 262144, "ps_before": [], "resident_before": false, "expires_at_before": null, "pinned_before": false, "keep_alive_policy": "bench instance: held for the arm's duration", "keep_alive_sent": "10m", "keep_alive_field_omitted": false, "think_mode": false, "think_note": "think:false accepted", "warmup_discarded": { "ttft_ms": 5695.26, "first_token_was_thinking": false, "wall_s": 7.3003, "eval_count": 256, "eval_duration_ns": 1602007000, "decode_tok_s": 159.8, "prompt_eval_count": 536, "prompt_eval_duration_ns": 258664000, "prefill_tok_s": 2072.186, "load_duration_ns": 5431597208, "total_duration_ns": 7296851274, "streamed_chunks": 256, "response_chars": 1072, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 158.878, "decode_tok_s_wall_gross": 35.067, "num_ctx_option": 131072 }, "ps_after_load": [ { "name": "gemma4:26b", "model": "gemma4:26b", "size": 17477150965, "digest": "5571076f3d70050487b26b341705799e0ab29b808164f90d20d4cf84f699d251", "details": { "parent_model": "", "format": "gguf", "family": "gemma4", "families": [ "gemma4" ], "parameter_size": "25.8B", "quantization_level": "Q4_K_M" }, "expires_at": "2026-09-21T21:21:13.740732627Z", "size_vram": 17477150965, "context_length": 131072 } ], "size_total": 17477150965, "size_vram": 17477150965, "context_length_loaded": 131072, "vram_after_load_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20711.0 }, "vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20696.0 }, "vram_receipt": "0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, 20711, 3274, 24463, 145.65, [N/A], 150.00, 43, N/A, [N/A], 2317, 14001, 94, 8, Active, Not Active, Not Active, Not Active", "fit_rule": "whole", "gpu": { "arm": "one", "requested_uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "fits_on_card": true, "size_bytes": 17477150965, "size_vram_bytes": 17477150965, "offload_fraction": 1.0, "offload_split": "100.0% VRAM / 0.0% RAM", "verdict": "fits: fully resident on the card", "smi_before": [ { "index": "0", "uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "memory.used": 15.0, "memory.free": 23970.0, "memory.total": 24463.0, "power.draw": 22.94, "power.limit": null, "enforced.power.limit": 150.0, "temperature.gpu": 35.0, "temperature.memory": null, "fan.speed": null, "clocks.sm": 1590.0, "clocks.mem": 14001.0, "utilization.gpu": 0.0, "pcie.link.width.current": 8.0, "clocks_throttle_reasons.sw_power_cap": "Not Active", "clocks_throttle_reasons.hw_slowdown": "Not Active", "clocks_throttle_reasons.sw_thermal_slowdown": "Not Active", "clocks_throttle_reasons.hw_thermal_slowdown": "Not Active" } ], "per_card_vram_delta_mib": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 20696.0 }, "smi_after": [ { "index": "0", "uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "memory.used": 20711.0, "memory.free": 3274.0, "memory.total": 24463.0, "power.draw": 146.12, "power.limit": null, "enforced.power.limit": 150.0, "temperature.gpu": 44.0, "temperature.memory": null, "fan.speed": null, "clocks.sm": 2250.0, "clocks.mem": 14001.0, "utilization.gpu": 94.0, "pcie.link.width.current": 8.0, "clocks_throttle_reasons.sw_power_cap": "Active", "clocks_throttle_reasons.hw_slowdown": "Not Active", "clocks_throttle_reasons.sw_thermal_slowdown": "Not Active", "clocks_throttle_reasons.hw_thermal_slowdown": "Not Active" } ] }, "scored": true, "contention_gates": [ { "run": 1, "contention_gate": { "attempts": [ { "window_s": 10.0, "samples": 20, "cards": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "utilization_pct": { "min": 0.0, "median": 0.0, "max": 94.0, "mean": 4.7, "n": 20 }, "power_w": { "min": 22.48, "median": 22.56, "max": 145.65, "mean": 33.24, "n": 20 }, "card_seen": true } }, "max_mean_util_pct": 4.7, "all_cards_seen": true, "attempt": 1, "verdict": "quiet: busiest card mean 4.7% <= 5.0%" } ], "passed": true, "cards": [ "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff" ] } }, { "run": 2, "contention_gate": { "attempts": [ { "window_s": 10.0, "samples": 20, "cards": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "utilization_pct": { "min": 0.0, "median": 0.0, "max": 93.0, "mean": 4.7, "n": 20 }, "power_w": { "min": 22.45, "median": 22.52, "max": 146.49, "mean": 32.3, "n": 20 }, "card_seen": true } }, "max_mean_util_pct": 4.7, "all_cards_seen": true, "attempt": 1, "verdict": "quiet: busiest card mean 4.7% <= 5.0%" } ], "passed": true, "cards": [ "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff" ] } }, { "run": 3, "contention_gate": { "attempts": [ { "window_s": 10.0, "samples": 20, "cards": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": { "utilization_pct": { "min": 0.0, "median": 0.0, "max": 94.0, "mean": 4.7, "n": 20 }, "power_w": { "min": 22.49, "median": 22.55, "max": 146.01, "mean": 32.3, "n": 20 }, "card_seen": true } }, "max_mean_util_pct": 4.7, "all_cards_seen": true, "attempt": 1, "verdict": "quiet: busiest card mean 4.7% <= 5.0%" } ], "passed": true, "cards": [ "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff" ] } } ], "peak_runner_rss_kb": 1745536, "peak_runner_rss_gib": 1.66, "peak_runner_threads": 31, "loadavg_during_runs": { "min": 1.87, "median": 2.12, "max": 2.28, "mean": 2.09, "samples": 88, "nproc": 24 }, "spread_gate": { "limit": 0.15, "values": [ 166.175, 165.656, 165.545 ], "spread_fraction": 0.0038, "verdict": "spread 0.4% of the median within the 15% gate" }, "voided_by_gate": false, "derived": { "decode_tok_s": { "min": 165.545, "median": 165.656, "max": 166.175, "mean": 165.792, "n": 3 }, "decode_tok_s_wall": { "min": 164.53, "median": 164.615, "max": 165.208, "mean": 164.784, "n": 3 }, "ttft_ms": { "min": 436.29, "median": 439.78, "max": 451.55, "mean": 442.54, "n": 3 }, "wall_s": { "min": 1.9862, "median": 1.9889, "max": 1.9951, "mean": 1.9901, "n": 3 }, "prefill_tok_s": { "min": 4599.832, "median": 4837.021, "max": 4940.502, "mean": 4792.452, "n": 3 }, "eval_count": { "min": 256, "median": 256, "max": 256, "mean": 256.0, "n": 3 }, "power_mean_w": { "min": 114.49, "median": 116.01, "max": 144.06, "mean": 124.85, "n": 3 }, "power_max_w": { "min": 146.73, "median": 147.54, "max": 147.7, "mean": 147.32, "n": 3 }, "power_over_idle_w": { "min": 92.0, "median": 93.53, "max": 121.59, "mean": 102.37, "n": 3 }, "energy_j_per_1k_tokens": { "min": 688.97, "median": 700.31, "max": 870.21, "mean": 753.16, "n": 3 }, "power_mean_w_all_cards": { "min": 114.49, "median": 116.01, "max": 144.06, "mean": 124.85, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 688.97, "median": 700.31, "max": 870.21, "mean": 753.16, "n": 3 }, "watts_per_tok_s": { "min": 0.689, "median": 0.7003, "max": 0.8702, "mean": 0.7532, "n": 3 }, "temp_max_c": { "min": 44.0, "median": 44.0, "max": 44.0, "mean": 44.0, "n": 3 }, "temp_rise_c": { "min": 1.0, "median": 2.0, "max": 2.0, "mean": 1.7, "n": 3 }, "fan_mean_pct": null, "fan_max_pct": null, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "hw_slowdown_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "clock_floor_mhz": { "min": 2107.0, "median": 2115.0, "max": 2122.0, "mean": 2114.7, "n": 3 }, "clock_max_mhz": { "min": 2265.0, "median": 2265.0, "max": 2287.0, "mean": 2272.3, "n": 3 }, "mem_clock_floor_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_median_mhz": { "min": 14001.0, "median": 14001.0, "max": 14001.0, "mean": 14001.0, "n": 3 }, "mem_clock_max_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": null, "decode_bandwidth_fraction_of_peak_pct": null, "thinking_chars_total": 0 }, "decode_bandwidth_refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here.", "memory_bus_width_bits": 256, "memory_bus_width_receipt": { "asked_utc": "2026-09-21T21:11:28Z", "method": "NVML nvmlDeviceGetMemoryBusWidth, per device, by UUID", "why": "nvidia-smi has no bus-width field, so without this the width would be a constant typed in from a specification rather than a reading.", "readable": true, "error": null, "widths_bits": { "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff": 256 } }, "bandwidth_formula_check": [ { "board": "GeForce RTX 3080", "mem_clock_mhz": 9501, "bus_bits": 320, "published_gbs": 760.3, "derived_gbs": 760.1, "error_pct": 0.026, "within_tolerance": true, "tolerance_pct": 0.5 }, { "board": "GeForce RTX 3080 Ti", "mem_clock_mhz": 9501, "bus_bits": 384, "published_gbs": 912.4, "derived_gbs": 912.1, "error_pct": 0.033, "within_tolerance": true, "tolerance_pct": 0.5 }, { "board": "GeForce RTX 3090", "mem_clock_mhz": 9751, "bus_bits": 384, "published_gbs": 936.2, "derived_gbs": 936.1, "error_pct": 0.011, "within_tolerance": true, "tolerance_pct": 0.5 }, { "board": "GeForce RTX 3090 Ti", "mem_clock_mhz": 10501, "bus_bits": 384, "published_gbs": 1008.0, "derived_gbs": 1008.1, "error_pct": 0.01, "within_tolerance": true, "tolerance_pct": 0.5 } ], "thermal_verdict": "run 1: NEITHER cap: the SM clock varied with draw at 76.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it; run 2: NEITHER cap: the SM clock varied with draw at 77.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it; run 3: the POWER CAP doing its job: the SM clock fell while draw sat at 96.0% of the 150 W cap, at 44 C -- well below any throttling temperature", "nvidia_after": "0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, 20711, 3274, 24463, 146.12, [N/A], 150.00, 44, N/A, [N/A], 2250, 14001, 94, 8, Active, Not Active, Not Active, Not Active", "smi_after": [ { "index": "0", "uuid": "GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "memory.used": 20711.0, "memory.free": 3274.0, "memory.total": 24463.0, "power.draw": 146.12, "power.limit": null, "enforced.power.limit": 150.0, "temperature.gpu": 44.0, "temperature.memory": null, "fan.speed": null, "clocks.sm": 2250.0, "clocks.mem": 14001.0, "utilization.gpu": 94.0, "pcie.link.width.current": 8.0, "clocks_throttle_reasons.sw_power_cap": "Active", "clocks_throttle_reasons.hw_slowdown": "Not Active", "clocks_throttle_reasons.sw_thermal_slowdown": "Not Active", "clocks_throttle_reasons.hw_thermal_slowdown": "Not Active" } ], "free_m_after": "total used free shared buff/cache available\nMem: 62698 28573 8737 4885 33553 34125\nSwap: 18431 18431 0", "ups_after_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-21T21:11:58Z", "available": false, "error": "no upsc on this box" }, "ps_final": [], "expires_at_after": null, "pinned_after": false, "finished_utc": "2026-09-21T21:11:58Z" }