{ "model_id": "m4", "model_tag": "mistral-small3.2:24b", "arm": "one", "box_class": "desktop-3090ti", "mode": "context-ladder", "arm_cards": [ "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4" ], "target_uuid": "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4", "started_utc": "2026-09-17T23:31:43Z", "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": "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 ollama 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: a partial residency is a different measurement and is never reported as the card's window.", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": "index 0 - EVGA - GPU-1809d1e1" }, "card_identity": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": 0, "bus_id": "00000000:01:00.0", "pci_sub_device_id": "0x49853842", "vendor": "EVGA", "board_per_operator": "EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB", "vbios": "94.02.A0.00.5C", "serial": null, "pcie_link_width_max": 16, "pcie_link_width_negotiated": 16, "memory_total_mib": 24564, "memory_bus_width_bits": 384, "clocks_max_sm_mhz": 2115, "clocks_max_memory_mhz": 10501, "power_default_limit_w": 450.0, "power_min_limit_w": 100.0, "power_max_limit_w": 480.0, "power_limit_at_bench_start_w": 300.0, "note": "the single card seated in this box's x16 slot -- the same slot the 3090 and the pair of 3080s were measured in. It came up at its own 450 W DEFAULT: the box's boot-time GPU power-cap unit (/etc/desktop/gpu-power-caps.conf, written 2026-08-27 for the two 3090s this box carried then) FAILED BOTH of its rows this boot -- neither UUID it names is in the machine any more -- and its own read-back printed `0, 00000000:01:00.0, 450.00 W, 450.00 W`. The operator then enabled persistence and set 300 W by hand, and the read-back is recorded per arm." } }, "memory_temp_support": { "checked_utc": "2026-09-17T23:31:43Z", "paths": { "nvidia-smi --query-gpu=temperature.memory": { "raw": "0, GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4, N/A", "readable": false }, "nvidia-smi -q -d TEMPERATURE": { "lines": [ "GPU Shutdown Temp : 97 C", "GPU Slowdown Temp : 94 C", "GPU Target Temperature : 83 C", "Memory Current Temp : N/A", "Memory Max Operating 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.", "instance_env": { "written_utc": "2026-09-17T23:29:11Z", "unit": "bench-3090ti-one", "shape": "one", "port": 11470, "base_url": "http://127.0.0.1:11470", "cuda_visible_devices": "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4", "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" }, "power_limits_at_start_w": "0, 450.00 W;", "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 16;" }, "ups_before_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-17T23:31:43Z", "available": false, "error": "upsc rc=1: Error: Connection failure: Connection refused" }, "levels": [ { "num_ctx": 4096, "started_utc": "2026-09-17T23:31:43Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15042.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 328.0 }, "runner_peak_rss_gib": 0.8, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 175284, "rss_kb": 840804, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2400 886 47 44347 44639\nSwap: 24575 795 23780", "kv_footprint_bytes": 0, "kv_footprint_gib": 0.0, "runs": [ { "ttft_ms": 308.01, "first_token_was_thinking": false, "wall_s": 4.7671, "eval_count": 256, "eval_duration_ns": 4457105000, "decode_tok_s": 57.436, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 39997000, "prefill_tok_s": 25576.918, "load_duration_ns": 264024918, "total_duration_ns": 4764933972, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.187, "decode_tok_s_wall_gross": 53.702, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 100.36, "median": 100.65, "max": 100.69, "mean": 100.57, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 100.57 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "power_w_all_cards": 100.57, "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": 15, "mean_w": 423.33, "max_w": 452.29, "median_w": 448.76, "mean_w_whole_window": 371.38, "max_w_whole_window": 452.29, "limit_w": 450.0, "over_idle_w": 322.76, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 100.6, "median": 448.51, "max": 452.29, "mean": 371.38, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.8, "n": 18 }, "temperature_c": { "min": 46.0, "median": 55.0, "max": 56.0, "mean": 53.7, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1950.0, "max": 1995.0, "mean": 1958.3, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 185.47, "median": 448.76, "max": 452.29, "mean": 423.33, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.1, "n": 15 }, "temperature_c": { "min": 54.0, "median": 55.0, "max": 56.0, "mean": 54.7, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1950.0, "max": 1995.0, "mean": 1962.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7370.42, "power_mean_w_all_cards": 423.33, "energy_j_per_1k_tokens_all_cards": 7370.42, "watts_per_tok_s": 7.3705, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 423.33, "power_pct_of_limit": 94.1, "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": false, "verdict": "no clock drop during decode (draw held at 94.1% of the 450 W cap) (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 423.33, "power_pct_of_limit": 94.1, "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": false, "verdict": "no clock drop during decode (draw held at 94.1% of the 450 W cap) (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.7, "fraction_of_peak_pct": 88.6, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 293.41, "first_token_was_thinking": false, "wall_s": 4.7543, "eval_count": 256, "eval_duration_ns": 4459097000, "decode_tok_s": 57.411, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 30426000, "prefill_tok_s": 33622.56, "load_duration_ns": 259074067, "total_duration_ns": 4752285335, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.163, "decode_tok_s_wall_gross": 53.846, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 17.72, "median": 17.81, "max": 17.96, "mean": 17.83, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17.83 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "power_w_all_cards": 17.83, "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": 16, "mean_w": 398.99, "max_w": 450.17, "median_w": 447.88, "mean_w_whole_window": 356.74, "max_w_whole_window": 450.17, "limit_w": 450.0, "over_idle_w": 381.16, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 17.84, "median": 447.42, "max": 450.17, "mean": 356.74, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 86.8, "n": 18 }, "temperature_c": { "min": 46.0, "median": 57.0, "max": 58.0, "mean": 55.4, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 48.0, "mean": 2.7, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 210.0, "median": 1965.0, "max": 1995.0, "mean": 1863.3, "n": 18 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 9704.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 93.47, "median": 447.88, "max": 450.17, "mean": 398.99, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 16 }, "utilization_pct": { "min": 93.0, "median": 98.0, "max": 99.0, "mean": 97.6, "n": 16 }, "temperature_c": { "min": 54.0, "median": 57.0, "max": 58.0, "mean": 56.6, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 48.0, "mean": 3.0, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1965.0, "max": 1995.0, "mean": 1963.1, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 6949.75, "power_mean_w_all_cards": 398.99, "energy_j_per_1k_tokens_all_cards": 6949.75, "watts_per_tok_s": 6.9497, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 58.0, "temp_min_c": 54.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 3.0, "fan_max_pct": 48.0, "clock_floor_mhz": 1920.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 398.99, "power_pct_of_limit": 88.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.875, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 58.0, "temp_min_c": 54.0, "temp_rise_c": 4.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 3.0, "fan_max_pct": 48.0, "clock_floor_mhz": 1920.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 398.99, "power_pct_of_limit": 88.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.875, "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": 984.1, "decode_gbs": 871.3, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 275.19, "first_token_was_thinking": false, "wall_s": 4.7367, "eval_count": 256, "eval_duration_ns": 4459638000, "decode_tok_s": 57.404, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20195000, "prefill_tok_s": 50656.103, "load_duration_ns": 251247866, "total_duration_ns": 4734709786, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.155, "decode_tok_s_wall_gross": 54.046, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 100.59, "median": 100.6, "max": 100.75, "mean": 100.65, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 100.65 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "power_w_all_cards": 100.65, "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": 15, "mean_w": 424.88, "max_w": 450.77, "median_w": 448.91, "mean_w_whole_window": 373.64, "max_w_whole_window": 450.77, "limit_w": 450.0, "over_idle_w": 324.23, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 100.74, "median": 448.78, "max": 450.77, "mean": 373.64, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 49.0, "median": 57.0, "max": 57.0, "mean": 55.9, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 69.0, "median": 69.0, "max": 70.0, "mean": 69.2, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 1995.0, "mean": 1957.5, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 199.13, "median": 448.91, "max": 450.77, "mean": 424.88, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15370.0, "median": 15370.0, "max": 15370.0, "mean": 15370.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.9, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 69.0, "median": 69.0, "max": 70.0, "mean": 69.1, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1965.0, "max": 1995.0, "mean": 1961.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7401.61, "power_mean_w_all_cards": 424.88, "energy_j_per_1k_tokens_all_cards": 7401.61, "watts_per_tok_s": 7.4016, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 69.1, "fan_max_pct": 70.0, "clock_floor_mhz": 1920.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 424.88, "power_pct_of_limit": 94.4, "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": false, "verdict": "no clock drop during decode (draw held at 94.4% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 69.1, "fan_max_pct": 70.0, "clock_floor_mhz": 1920.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 424.88, "power_pct_of_limit": 94.4, "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": false, "verdict": "no clock drop during decode (draw held at 94.4% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.2, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.404, "median": 57.411, "max": 57.436, "mean": 57.417, "n": 3 }, "ttft_ms": { "min": 275.19, "median": 293.41, "max": 308.01, "mean": 292.2, "n": 3 }, "power_mean_w": { "min": 398.99, "median": 423.33, "max": 424.88, "mean": 415.73, "n": 3 }, "power_mean_w_all_cards": { "min": 398.99, "median": 423.33, "max": 424.88, "mean": 415.73, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 6949.75, "median": 7370.42, "max": 7401.61, "mean": 7240.59, "n": 3 }, "energy_j_per_1k_tokens": { "min": 6949.75, "median": 7370.42, "max": 7401.61, "mean": 7240.59, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 871.2, "median": 871.3, "max": 871.7, "mean": 871.4, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.5, "median": 88.5, "max": 88.6, "mean": 88.5, "n": 3 }, "sw_power_cap_active_fraction": { "min": 0.875, "median": 1.0, "max": 1.0, "mean": 0.958, "n": 3 }, "temp_max_c": { "min": 56.0, "median": 57.0, "max": 58.0, "mean": 57.0, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 48.0, "max": 70.0, "mean": 39.3, "n": 3 } }, "finished_utc": "2026-09-17T23:32:58Z" }, { "num_ctx": 8192, "started_utc": "2026-09-17T23:32:58Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 588.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15370.0 }, "runner_peak_rss_gib": 0.84, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 176291, "rss_kb": 877144, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2481 717 79 44467 44558\nSwap: 24575 795 23780", "kv_footprint_bytes": 599795958, "kv_footprint_gib": 0.56, "runs": [ { "ttft_ms": 288.51, "first_token_was_thinking": false, "wall_s": 4.7525, "eval_count": 256, "eval_duration_ns": 4462119000, "decode_tok_s": 57.372, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 27671000, "prefill_tok_s": 36970.113, "load_duration_ns": 257903527, "total_duration_ns": 4750480629, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.124, "decode_tok_s_wall_gross": 53.867, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 98.89, "median": 100.13, "max": 101.4, "mean": 100.14, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 100.14 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "power_w_all_cards": 100.14, "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": 15, "mean_w": 423.61, "max_w": 449.79, "median_w": 448.07, "mean_w_whole_window": 372.26, "max_w_whole_window": 449.79, "limit_w": 450.0, "over_idle_w": 323.47, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 101.27, "median": 447.76, "max": 449.79, "mean": 372.26, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 47.0, "median": 55.0, "max": 56.0, "mean": 54.1, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 63.0, "median": 64.0, "max": 67.0, "mean": 64.5, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1965.0, "max": 1995.0, "mean": 1963.3, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 191.98, "median": 448.07, "max": 449.79, "mean": 423.61, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 54.0, "median": 55.0, "max": 56.0, "mean": 55.1, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 63.0, "median": 64.0, "max": 66.0, "mean": 64.1, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1968.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7383.59, "power_mean_w_all_cards": 423.61, "energy_j_per_1k_tokens_all_cards": 7383.59, "watts_per_tok_s": 7.3836, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 64.1, "fan_max_pct": 66.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 423.61, "power_pct_of_limit": 94.1, "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": false, "verdict": "no clock drop during decode (draw held at 94.1% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 64.1, "fan_max_pct": 66.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 423.61, "power_pct_of_limit": 94.1, "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": false, "verdict": "no clock drop during decode (draw held at 94.1% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.8, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 265.69, "first_token_was_thinking": false, "wall_s": 4.7323, "eval_count": 256, "eval_duration_ns": 4464812000, "decode_tok_s": 57.337, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20490000, "prefill_tok_s": 49926.794, "load_duration_ns": 242212784, "total_duration_ns": 4730244899, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.091, "decode_tok_s_wall_gross": 54.097, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 101.41, "median": 101.54, "max": 101.7, "mean": 101.55, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 101.55 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "power_w_all_cards": 101.55, "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": 408.87, "max_w": 450.38, "median_w": 448.85, "mean_w_whole_window": 374.73, "max_w_whole_window": 450.38, "limit_w": 450.0, "over_idle_w": 307.32, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 101.51, "median": 448.51, "max": 450.38, "mean": 374.73, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 87.0, "n": 18 }, "temperature_c": { "min": 47.0, "median": 55.0, "max": 56.0, "mean": 54.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 52.0, "mean": 49.5, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1965.0, "max": 1995.0, "mean": 1960.8, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 156.64, "median": 448.85, "max": 450.38, "mean": 408.87, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 16 }, "utilization_pct": { "min": 91.0, "median": 99.0, "max": 99.0, "mean": 97.9, "n": 16 }, "temperature_c": { "min": 54.0, "median": 55.0, "max": 56.0, "mean": 55.2, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 51.0, "mean": 49.2, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1972.5, "max": 1995.0, "mean": 1965.9, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 7130.97, "power_mean_w_all_cards": 408.87, "energy_j_per_1k_tokens_all_cards": 7130.97, "watts_per_tok_s": 7.131, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 49.2, "fan_max_pct": 51.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1972.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.87, "power_pct_of_limit": 90.9, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.9% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 49.2, "fan_max_pct": 51.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1972.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.87, "power_pct_of_limit": 90.9, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.9% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.2, "fraction_of_peak_pct": 88.4, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 277.21, "first_token_was_thinking": false, "wall_s": 4.741, "eval_count": 256, "eval_duration_ns": 4461942000, "decode_tok_s": 57.374, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 29480000, "prefill_tok_s": 34701.493, "load_duration_ns": 244502120, "total_duration_ns": 4738906752, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.127, "decode_tok_s_wall_gross": 53.998, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 17.42, "median": 17.45, "max": 17.67, "mean": 17.51, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17.51 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "power_w_all_cards": 17.51, "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": 15, "mean_w": 419.82, "max_w": 449.88, "median_w": 447.96, "mean_w_whole_window": 357.25, "max_w_whole_window": 449.88, "limit_w": 450.0, "over_idle_w": 402.31, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 17.64, "median": 447.59, "max": 449.88, "mean": 357.25, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 46.0, "median": 56.5, "max": 57.0, "mean": 55.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 47.0, "mean": 2.6, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 210.0, "median": 1965.0, "max": 1995.0, "mean": 1865.0, "n": 18 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 9704.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 136.65, "median": 447.96, "max": 449.88, "mean": 419.82, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 15958.0, "median": 15958.0, "max": 15958.0, "mean": 15958.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.6, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 47.0, "mean": 3.1, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1964.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7317.24, "power_mean_w_all_cards": 419.82, "energy_j_per_1k_tokens_all_cards": 7317.24, "watts_per_tok_s": 7.3173, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 3.1, "fan_max_pct": 47.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 419.82, "power_pct_of_limit": 93.3, "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 93.3% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 3.1, "fan_max_pct": 47.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 419.82, "power_pct_of_limit": 93.3, "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 93.3% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.8, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.337, "median": 57.372, "max": 57.374, "mean": 57.361, "n": 3 }, "ttft_ms": { "min": 265.69, "median": 277.21, "max": 288.51, "mean": 277.14, "n": 3 }, "power_mean_w": { "min": 408.87, "median": 419.82, "max": 423.61, "mean": 417.43, "n": 3 }, "power_mean_w_all_cards": { "min": 408.87, "median": 419.82, "max": 423.61, "mean": 417.43, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 7130.97, "median": 7317.24, "max": 7383.59, "mean": 7277.27, "n": 3 }, "energy_j_per_1k_tokens": { "min": 7130.97, "median": 7317.24, "max": 7383.59, "mean": 7277.27, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 870.2, "median": 870.8, "max": 870.8, "mean": 870.6, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.4, "median": 88.5, "max": 88.5, "mean": 88.5, "n": 3 }, "sw_power_cap_active_fraction": { "min": 0.875, "median": 1.0, "max": 1.0, "mean": 0.958, "n": 3 }, "temp_max_c": { "min": 56.0, "median": 56.0, "max": 57.0, "mean": 56.3, "n": 3 }, "fan_max_pct": { "min": 47.0, "median": 51.0, "max": 66.0, "mean": 54.7, "n": 3 } }, "finished_utc": "2026-09-17T23:34:06Z" }, { "num_ctx": 16384, "started_utc": "2026-09-17T23:34:06Z", "loads": true, "size_total": 15685124095, "size_vram": 15685124095, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 444.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 15958.0 }, "runner_peak_rss_gib": 0.82, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 177440, "rss_kb": 855060, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2475 720 59 44450 44564\nSwap: 24575 795 23780", "kv_footprint_bytes": 1082130432, "kv_footprint_gib": 1.01, "runs": [ { "ttft_ms": 279.77, "first_token_was_thinking": false, "wall_s": 4.7419, "eval_count": 256, "eval_duration_ns": 4460176000, "decode_tok_s": 57.397, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28263000, "prefill_tok_s": 36195.733, "load_duration_ns": 248370479, "total_duration_ns": 4739746100, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.148, "decode_tok_s_wall_gross": 53.987, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 101.28, "median": 101.79, "max": 102.48, "mean": 101.85, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 101.85 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "power_w_all_cards": 101.85, "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": 16, "mean_w": 408.0, "max_w": 450.61, "median_w": 448.62, "mean_w_whole_window": 374.07, "max_w_whole_window": 450.61, "limit_w": 450.0, "over_idle_w": 306.15, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.61, "median": 448.18, "max": 450.61, "mean": 374.07, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 87.2, "n": 18 }, "temperature_c": { "min": 49.0, "median": 56.0, "max": 57.0, "mean": 55.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 59.0, "median": 59.0, "max": 62.0, "mean": 59.7, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 1995.0, "mean": 1955.0, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 152.51, "median": 448.62, "max": 450.61, "mean": 408.0, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 16 }, "utilization_pct": { "min": 97.0, "median": 98.0, "max": 99.0, "mean": 98.1, "n": 16 }, "temperature_c": { "min": 55.0, "median": 56.5, "max": 57.0, "mean": 56.4, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 59.0, "median": 59.0, "max": 61.0, "mean": 59.5, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1959.4, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 7108.41, "power_mean_w_all_cards": 408.0, "energy_j_per_1k_tokens_all_cards": 7108.41, "watts_per_tok_s": 7.1084, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 59.5, "fan_max_pct": 61.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.0, "power_pct_of_limit": 90.7, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.7% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 59.5, "fan_max_pct": 61.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.0, "power_pct_of_limit": 90.7, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.7% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.1, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 278.82, "first_token_was_thinking": false, "wall_s": 4.7373, "eval_count": 256, "eval_duration_ns": 4456706000, "decode_tok_s": 57.442, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20393000, "prefill_tok_s": 50164.272, "load_duration_ns": 255474178, "total_duration_ns": 4735357143, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.195, "decode_tok_s_wall_gross": 54.039, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.12, "median": 102.35, "max": 102.36, "mean": 102.28, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.28 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "power_w_all_cards": 102.28, "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": 15, "mean_w": 427.05, "max_w": 452.45, "median_w": 449.01, "mean_w_whole_window": 375.6, "max_w_whole_window": 452.45, "limit_w": 450.0, "over_idle_w": 324.77, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.38, "median": 448.95, "max": 452.45, "mean": 375.6, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.2, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 56.0, "median": 57.0, "max": 60.0, "mean": 57.6, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 1995.0, "mean": 1956.7, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 199.74, "median": 449.01, "max": 452.45, "mean": 427.05, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 57.0, "mean": 56.1, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 56.0, "median": 57.0, "max": 59.0, "mean": 57.3, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1960.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7434.52, "power_mean_w_all_cards": 427.05, "energy_j_per_1k_tokens_all_cards": 7434.52, "watts_per_tok_s": 7.4345, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 57.3, "fan_max_pct": 59.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 427.05, "power_pct_of_limit": 94.9, "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 94.9% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 57.3, "fan_max_pct": 59.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 427.05, "power_pct_of_limit": 94.9, "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 94.9% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.8, "fraction_of_peak_pct": 88.6, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 277.96, "first_token_was_thinking": false, "wall_s": 4.7367, "eval_count": 256, "eval_duration_ns": 4456941000, "decode_tok_s": 57.438, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20317000, "prefill_tok_s": 50351.922, "load_duration_ns": 254652904, "total_duration_ns": 4734547430, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.191, "decode_tok_s_wall_gross": 54.046, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.54, "median": 102.54, "max": 102.76, "mean": 102.61, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.61 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "power_w_all_cards": 102.61, "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": 14, "mean_w": 441.65, "max_w": 451.24, "median_w": 448.94, "mean_w_whole_window": 375.02, "max_w_whole_window": 451.24, "limit_w": 450.0, "over_idle_w": 339.04, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.52, "median": 448.48, "max": 451.24, "mean": 375.02, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 76.4, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.4, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 49.0, "median": 50.0, "max": 52.0, "mean": 50.6, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1950.0, "max": 1995.0, "mean": 1954.2, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8889, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 344.4, "median": 448.94, "max": 451.24, "mean": 441.65, "n": 14 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 16402.0, "median": 16402.0, "max": 16402.0, "mean": 16402.0, "n": 14 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 14 }, "temperature_c": { "min": 56.0, "median": 56.5, "max": 57.0, "mean": 56.5, "n": 14 }, "memory_temperature_c": null, "fan_pct": { "min": 49.0, "median": 50.0, "max": 51.0, "mean": 50.1, "n": 14 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 14 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1950.0, "max": 1980.0, "mean": 1954.3, "n": 14 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 14 }, "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": 14 }, "busy_samples": 14 } } }, "energy_j_per_1k_tokens": 7689.09, "power_mean_w_all_cards": 441.65, "energy_j_per_1k_tokens_all_cards": 7689.09, "watts_per_tok_s": 7.6892, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 14, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 50.1, "fan_max_pct": 51.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1980.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 441.65, "power_pct_of_limit": 98.1, "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": false, "verdict": "no clock drop during decode (draw held at 98.1% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 14, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 50.1, "fan_max_pct": 51.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1980.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 441.65, "power_pct_of_limit": 98.1, "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": false, "verdict": "no clock drop during decode (draw held at 98.1% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.8, "fraction_of_peak_pct": 88.6, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.397, "median": 57.438, "max": 57.442, "mean": 57.426, "n": 3 }, "ttft_ms": { "min": 277.96, "median": 278.82, "max": 279.77, "mean": 278.85, "n": 3 }, "power_mean_w": { "min": 408.0, "median": 427.05, "max": 441.65, "mean": 425.57, "n": 3 }, "power_mean_w_all_cards": { "min": 408.0, "median": 427.05, "max": 441.65, "mean": 425.57, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 7108.41, "median": 7434.52, "max": 7689.09, "mean": 7410.67, "n": 3 }, "energy_j_per_1k_tokens": { "min": 7108.41, "median": 7434.52, "max": 7689.09, "mean": 7410.67, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 871.1, "median": 871.8, "max": 871.8, "mean": 871.6, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.5, "median": 88.6, "max": 88.6, "mean": 88.6, "n": 3 }, "sw_power_cap_active_fraction": { "min": 0.875, "median": 1.0, "max": 1.0, "mean": 0.958, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 3 }, "fan_max_pct": { "min": 51.0, "median": 59.0, "max": 61.0, "mean": 57.0, "n": 3 } }, "finished_utc": "2026-09-17T23:35:03Z" }, { "num_ctx": 32768, "started_utc": "2026-09-17T23:35:03Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 1376.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16402.0 }, "runner_peak_rss_gib": 0.83, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 178111, "rss_kb": 872136, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2541 654 75 44466 44499\nSwap: 24575 795 23780", "kv_footprint_bytes": 2524971008, "kv_footprint_gib": 2.35, "runs": [ { "ttft_ms": 277.87, "first_token_was_thinking": false, "wall_s": 4.7521, "eval_count": 256, "eval_duration_ns": 4472345000, "decode_tok_s": 57.241, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28510000, "prefill_tok_s": 35882.147, "load_duration_ns": 246363080, "total_duration_ns": 4750018259, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 56.993, "decode_tok_s_wall_gross": 53.871, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 101.49, "median": 101.65, "max": 101.68, "mean": 101.61, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 101.61 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "power_w_all_cards": 101.61, "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": 408.2, "max_w": 451.2, "median_w": 448.29, "mean_w_whole_window": 374.23, "max_w_whole_window": 451.2, "limit_w": 450.0, "over_idle_w": 306.59, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.2, "median": 448.2, "max": 451.2, "mean": 374.23, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 87.2, "n": 18 }, "temperature_c": { "min": 49.0, "median": 56.0, "max": 57.0, "mean": 55.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 53.0, "max": 55.0, "mean": 53.1, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1950.0, "max": 1995.0, "mean": 1953.3, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 150.27, "median": 448.29, "max": 451.2, "mean": 408.2, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 16 }, "utilization_pct": { "min": 97.0, "median": 98.0, "max": 99.0, "mean": 98.1, "n": 16 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 57.0, "mean": 56.4, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 53.0, "max": 54.0, "mean": 52.9, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1957.5, "max": 1995.0, "mean": 1957.5, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 7131.29, "power_mean_w_all_cards": 408.2, "energy_j_per_1k_tokens_all_cards": 7131.29, "watts_per_tok_s": 7.1313, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 52.9, "fan_max_pct": 54.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1957.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.2, "power_pct_of_limit": 90.7, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.7% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 52.9, "fan_max_pct": 54.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1957.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.2, "power_pct_of_limit": 90.7, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.7% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 868.8, "fraction_of_peak_pct": 88.3, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 268.89, "first_token_was_thinking": false, "wall_s": 4.7399, "eval_count": 256, "eval_duration_ns": 4469227000, "decode_tok_s": 57.281, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20304000, "prefill_tok_s": 50384.161, "load_duration_ns": 245368022, "total_duration_ns": 4737907160, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.034, "decode_tok_s_wall_gross": 54.01, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.68, "median": 99.85, "max": 99.92, "mean": 99.82, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.82 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "power_w_all_cards": 99.82, "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": 15, "mean_w": 426.08, "max_w": 451.82, "median_w": 448.96, "mean_w_whole_window": 375.18, "max_w_whole_window": 451.82, "limit_w": 450.0, "over_idle_w": 326.26, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.67, "median": 448.9, "max": 451.82, "mean": 375.18, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 82.0, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 58.0, "median": 59.0, "max": 61.0, "mean": 59.2, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 1995.0, "mean": 1957.5, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 194.4, "median": 448.96, "max": 451.82, "mean": 426.08, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.4, "n": 15 }, "temperature_c": { "min": 56.0, "median": 56.0, "max": 57.0, "mean": 56.3, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 58.0, "median": 59.0, "max": 60.0, "mean": 58.8, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1960.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7438.47, "power_mean_w_all_cards": 426.08, "energy_j_per_1k_tokens_all_cards": 7438.47, "watts_per_tok_s": 7.4384, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 58.8, "fan_max_pct": 60.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 426.08, "power_pct_of_limit": 94.7, "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 94.7% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 58.8, "fan_max_pct": 60.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 426.08, "power_pct_of_limit": 94.7, "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 94.7% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 869.4, "fraction_of_peak_pct": 88.3, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 263.67, "first_token_was_thinking": false, "wall_s": 4.7342, "eval_count": 256, "eval_duration_ns": 4468784000, "decode_tok_s": 57.286, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20142000, "prefill_tok_s": 50789.395, "load_duration_ns": 240570029, "total_duration_ns": 4732074336, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.041, "decode_tok_s_wall_gross": 54.075, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.88, "median": 99.92, "max": 100.07, "mean": 99.96, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.96 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "power_w_all_cards": 99.96, "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": 410.48, "max_w": 452.11, "median_w": 448.56, "mean_w_whole_window": 375.99, "max_w_whole_window": 452.11, "limit_w": 450.0, "over_idle_w": 310.52, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.45, "median": 448.22, "max": 452.11, "mean": 375.99, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 87.1, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.3, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 53.0, "median": 53.5, "max": 56.0, "mean": 53.9, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 1995.0, "mean": 1956.7, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 152.62, "median": 448.56, "max": 452.11, "mean": 410.48, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 17778.0, "median": 17778.0, "max": 17778.0, "mean": 17778.0, "n": 16 }, "utilization_pct": { "min": 97.0, "median": 98.0, "max": 99.0, "mean": 98.0, "n": 16 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 57.0, "mean": 56.2, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 53.0, "median": 53.0, "max": 55.0, "mean": 53.6, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 1995.0, "mean": 1961.2, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 7165.42, "power_mean_w_all_cards": 410.48, "energy_j_per_1k_tokens_all_cards": 7165.42, "watts_per_tok_s": 7.1655, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 53.6, "fan_max_pct": 55.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 410.48, "power_pct_of_limit": 91.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 91.2% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 53.6, "fan_max_pct": 55.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 410.48, "power_pct_of_limit": 91.2, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 91.2% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 869.5, "fraction_of_peak_pct": 88.4, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.241, "median": 57.281, "max": 57.286, "mean": 57.269, "n": 3 }, "ttft_ms": { "min": 263.67, "median": 268.89, "max": 277.87, "mean": 270.14, "n": 3 }, "power_mean_w": { "min": 408.2, "median": 410.48, "max": 426.08, "mean": 414.92, "n": 3 }, "power_mean_w_all_cards": { "min": 408.2, "median": 410.48, "max": 426.08, "mean": 414.92, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 7131.29, "median": 7165.42, "max": 7438.47, "mean": 7245.06, "n": 3 }, "energy_j_per_1k_tokens": { "min": 7131.29, "median": 7165.42, "max": 7438.47, "mean": 7245.06, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 868.8, "median": 869.4, "max": 869.5, "mean": 869.2, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.3, "median": 88.3, "max": 88.4, "mean": 88.3, "n": 3 }, "sw_power_cap_active_fraction": { "min": 0.875, "median": 0.875, "max": 1.0, "mean": 0.917, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 3 }, "fan_max_pct": { "min": 54.0, "median": 55.0, "max": 60.0, "mean": 56.3, "n": 3 } }, "finished_utc": "2026-09-17T23:36:01Z" }, { "num_ctx": 49152, "started_utc": "2026-09-17T23:36:01Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 1384.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17778.0 }, "runner_peak_rss_gib": 0.85, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 178774, "rss_kb": 889204, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2506 689 91 44483 44534\nSwap: 24575 795 23780", "kv_footprint_bytes": 3976200192, "kv_footprint_gib": 3.7, "runs": [ { "ttft_ms": 275.81, "first_token_was_thinking": false, "wall_s": 4.7416, "eval_count": 256, "eval_duration_ns": 4463996000, "decode_tok_s": 57.348, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28168000, "prefill_tok_s": 36317.807, "load_duration_ns": 244481293, "total_duration_ns": 4739449224, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.1, "decode_tok_s_wall_gross": 53.99, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.52, "median": 99.66, "max": 99.84, "mean": 99.67, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.67 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "power_w_all_cards": 99.67, "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": 15, "mean_w": 425.67, "max_w": 451.39, "median_w": 448.52, "mean_w_whole_window": 374.4, "max_w_whole_window": 451.39, "limit_w": 450.0, "over_idle_w": 326.0, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.85, "median": 448.27, "max": 451.39, "mean": 374.4, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 51.0, "mean": 48.8, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 2010.0, "mean": 1958.3, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 203.54, "median": 448.52, "max": 451.39, "mean": 425.67, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.5, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 48.0, "max": 50.0, "mean": 48.5, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 2010.0, "mean": 1961.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7422.61, "power_mean_w_all_cards": 425.67, "energy_j_per_1k_tokens_all_cards": 7422.61, "watts_per_tok_s": 7.4226, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.5, "fan_max_pct": 50.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 425.67, "power_pct_of_limit": 94.6, "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 94.6% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.5, "fan_max_pct": 50.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 425.67, "power_pct_of_limit": 94.6, "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 94.6% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.4, "fraction_of_peak_pct": 88.4, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 270.9, "first_token_was_thinking": false, "wall_s": 4.7363, "eval_count": 256, "eval_duration_ns": 4463565000, "decode_tok_s": 57.353, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20310000, "prefill_tok_s": 50369.276, "load_duration_ns": 247441652, "total_duration_ns": 4734293690, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.105, "decode_tok_s_wall_gross": 54.05, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.52, "median": 99.71, "max": 99.77, "mean": 99.67, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.67 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "power_w_all_cards": 99.67, "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": 12, "mean_w": 448.87, "max_w": 452.69, "median_w": 448.74, "mean_w_whole_window": 374.87, "max_w_whole_window": 452.69, "limit_w": 450.0, "over_idle_w": 349.2, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.73, "median": 448.31, "max": 452.69, "mean": 374.87, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 65.6, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 55.4, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 61.0, "median": 61.0, "max": 64.0, "mean": 61.7, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1950.0, "max": 2010.0, "mean": 1959.2, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 446.9, "median": 448.74, "max": 452.69, "mean": 448.87, "n": 12 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 12 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 12 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.6, "n": 12 }, "memory_temperature_c": null, "fan_pct": { "min": 61.0, "median": 61.0, "max": 62.0, "mean": 61.1, "n": 12 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 12 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1950.0, "max": 1980.0, "mean": 1957.5, "n": 12 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 12 }, "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": 12 }, "busy_samples": 12 } } }, "energy_j_per_1k_tokens": 7826.41, "power_mean_w_all_cards": 448.87, "energy_j_per_1k_tokens_all_cards": 7826.41, "watts_per_tok_s": 7.8264, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 12, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 61.1, "fan_max_pct": 62.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1980.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 448.87, "power_pct_of_limit": 99.7, "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 99.7% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 12, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 61.1, "fan_max_pct": 62.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 1980.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 448.87, "power_pct_of_limit": 99.7, "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 99.7% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.5, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 272.21, "first_token_was_thinking": false, "wall_s": 4.7361, "eval_count": 256, "eval_duration_ns": 4461866000, "decode_tok_s": 57.375, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20124000, "prefill_tok_s": 50834.824, "load_duration_ns": 249163889, "total_duration_ns": 4734061217, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.125, "decode_tok_s_wall_gross": 54.053, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 98.85, "median": 99.06, "max": 99.15, "mean": 99.02, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.02 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "power_w_all_cards": 99.02, "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": 15, "mean_w": 425.83, "max_w": 452.61, "median_w": 448.92, "mean_w_whole_window": 374.39, "max_w_whole_window": 452.61, "limit_w": 450.0, "over_idle_w": 326.81, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 98.78, "median": 448.62, "max": 452.61, "mean": 374.39, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.8, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.0, "max": 57.0, "mean": 54.9, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 56.0, "median": 57.0, "max": 59.0, "mean": 57.1, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1950.0, "max": 2010.0, "mean": 1959.2, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 202.05, "median": 448.92, "max": 452.61, "mean": 425.83, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 19162.0, "median": 19162.0, "max": 19162.0, "mean": 19162.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.1, "n": 15 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 57.0, "mean": 55.9, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 56.0, "median": 56.0, "max": 58.0, "mean": 56.7, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1950.0, "max": 2010.0, "mean": 1963.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7421.86, "power_mean_w_all_cards": 425.83, "energy_j_per_1k_tokens_all_cards": 7421.86, "watts_per_tok_s": 7.4219, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 56.7, "fan_max_pct": 58.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 425.83, "power_pct_of_limit": 94.6, "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 94.6% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 56.7, "fan_max_pct": 58.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1950.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 425.83, "power_pct_of_limit": 94.6, "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 94.6% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.8, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.348, "median": 57.353, "max": 57.375, "mean": 57.359, "n": 3 }, "ttft_ms": { "min": 270.9, "median": 272.21, "max": 275.81, "mean": 272.97, "n": 3 }, "power_mean_w": { "min": 425.67, "median": 425.83, "max": 448.87, "mean": 433.46, "n": 3 }, "power_mean_w_all_cards": { "min": 425.67, "median": 425.83, "max": 448.87, "mean": 433.46, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 7421.86, "median": 7422.61, "max": 7826.41, "mean": 7556.96, "n": 3 }, "energy_j_per_1k_tokens": { "min": 7421.86, "median": 7422.61, "max": 7826.41, "mean": 7556.96, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 870.4, "median": 870.5, "max": 870.8, "mean": 870.6, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.4, "median": 88.5, "max": 88.5, "mean": 88.5, "n": 3 }, "sw_power_cap_active_fraction": { "min": 1.0, "median": 1.0, "max": 1.0, "mean": 1.0, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 3 }, "fan_max_pct": { "min": 50.0, "median": 58.0, "max": 62.0, "mean": 56.7, "n": 3 } }, "finished_utc": "2026-09-17T23:36:59Z" }, { "num_ctx": 65536, "started_utc": "2026-09-17T23:36:59Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 1440.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19162.0 }, "runner_peak_rss_gib": 0.86, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 179587, "rss_kb": 906452, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2546 648 107 44499 44494\nSwap: 24575 795 23780", "kv_footprint_bytes": 5486149632, "kv_footprint_gib": 5.11, "runs": [ { "ttft_ms": 275.63, "first_token_was_thinking": false, "wall_s": 4.7376, "eval_count": 256, "eval_duration_ns": 4460223000, "decode_tok_s": 57.396, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28318000, "prefill_tok_s": 36125.433, "load_duration_ns": 244375674, "total_duration_ns": 4735672181, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.15, "decode_tok_s_wall_gross": 54.036, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.24, "median": 99.32, "max": 99.42, "mean": 99.33, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.33 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "power_w_all_cards": 99.33, "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": 15, "mean_w": 426.1, "max_w": 451.3, "median_w": 448.85, "mean_w_whole_window": 375.01, "max_w_whole_window": 451.3, "limit_w": 450.0, "over_idle_w": 326.77, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.15, "median": 448.8, "max": 451.3, "mean": 375.01, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.8, "n": 18 }, "temperature_c": { "min": 48.0, "median": 56.5, "max": 57.0, "mean": 55.7, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 45.0, "median": 45.5, "max": 48.0, "mean": 45.8, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 2010.0, "mean": 1959.2, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 194.35, "median": 448.85, "max": 451.3, "mean": 426.1, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.1, "n": 15 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.6, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 45.0, "median": 45.0, "max": 47.0, "mean": 45.5, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 2010.0, "mean": 1962.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7423.83, "power_mean_w_all_cards": 426.1, "energy_j_per_1k_tokens_all_cards": 7423.83, "watts_per_tok_s": 7.4239, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 45.5, "fan_max_pct": 47.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 426.1, "power_pct_of_limit": 94.7, "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 94.7% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 45.5, "fan_max_pct": 47.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 426.1, "power_pct_of_limit": 94.7, "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 94.7% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.1, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 271.37, "first_token_was_thinking": false, "wall_s": 4.7356, "eval_count": 256, "eval_duration_ns": 4462398000, "decode_tok_s": 57.368, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 20518000, "prefill_tok_s": 49858.661, "load_duration_ns": 247706788, "total_duration_ns": 4733434120, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.12, "decode_tok_s_wall_gross": 54.058, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 100.06, "median": 100.28, "max": 100.53, "mean": 100.29, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 100.29 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "power_w_all_cards": 100.29, "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": 408.83, "max_w": 452.8, "median_w": 448.62, "mean_w_whole_window": 374.58, "max_w_whole_window": 452.8, "limit_w": 450.0, "over_idle_w": 308.54, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.99, "median": 448.3, "max": 452.8, "mean": 374.58, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 86.4, "n": 18 }, "temperature_c": { "min": 49.0, "median": 56.0, "max": 57.0, "mean": 55.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 63.0, "median": 64.0, "max": 66.0, "mean": 64.1, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 1920.0, "median": 1957.5, "max": 2010.0, "mean": 1959.2, "n": 18 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.7778, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 149.95, "median": 448.62, "max": 452.8, "mean": 408.83, "n": 16 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 16 }, "utilization_pct": { "min": 91.0, "median": 98.0, "max": 99.0, "mean": 97.2, "n": 16 }, "temperature_c": { "min": 56.0, "median": 56.0, "max": 57.0, "mean": 56.4, "n": 16 }, "memory_temperature_c": null, "fan_pct": { "min": 63.0, "median": 63.5, "max": 66.0, "mean": 63.8, "n": 16 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 16 }, "clocks_sm_mhz": { "min": 1935.0, "median": 1965.0, "max": 2010.0, "mean": 1964.1, "n": 16 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 16 }, "throttle_sw_power_cap_fraction": 0.875, "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": 7126.41, "power_mean_w_all_cards": 408.83, "energy_j_per_1k_tokens_all_cards": 7126.41, "watts_per_tok_s": 7.1264, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 63.8, "fan_max_pct": 66.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.83, "power_pct_of_limit": 90.9, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.9% of the 450 W cap)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 16, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 63.8, "fan_max_pct": 66.0, "clock_floor_mhz": 1935.0, "clock_max_mhz": 2010.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 408.83, "power_pct_of_limit": 90.9, "power_pinned_at_limit": true, "sw_power_cap_active_fraction": 0.875, "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 90.9% of the 450 W cap)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 870.7, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 284.3, "first_token_was_thinking": false, "wall_s": 4.7461, "eval_count": 256, "eval_duration_ns": 4459941000, "decode_tok_s": 57.4, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 29491000, "prefill_tok_s": 34688.549, "load_duration_ns": 251691924, "total_duration_ns": 4744116511, "streamed_chunks": 256, "response_chars": 1019, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 57.152, "decode_tok_s_wall_gross": 53.939, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 16.78, "median": 16.94, "max": 17.08, "mean": 16.93, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16.93 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "power_w_all_cards": 16.93, "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": 15, "mean_w": 418.76, "max_w": 450.0, "median_w": 448.09, "mean_w_whole_window": 355.94, "max_w_whole_window": 450.0, "limit_w": 450.0, "over_idle_w": 401.83, "samples": 18, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 17.07, "median": 447.74, "max": 450.0, "mean": 355.94, "n": 18 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 18 }, "utilization_pct": { "min": 0.0, "median": 98.0, "max": 99.0, "mean": 81.9, "n": 18 }, "temperature_c": { "min": 44.0, "median": 55.0, "max": 56.0, "mean": 53.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 41.0, "mean": 6.7, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 210.0, "median": 1965.0, "max": 1995.0, "mean": 1870.8, "n": 18 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 9704.0, "n": 18 }, "throttle_sw_power_cap_fraction": 0.8333, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 130.56, "median": 448.09, "max": 450.0, "mean": 418.76, "n": 15 }, "power_limit_w": 450.0, "memory_used_mib": { "min": 20602.0, "median": 20602.0, "max": 20602.0, "mean": 20602.0, "n": 15 }, "utilization_pct": { "min": 98.0, "median": 98.0, "max": 99.0, "mean": 98.3, "n": 15 }, "temperature_c": { "min": 54.0, "median": 55.0, "max": 56.0, "mean": 54.9, "n": 15 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 15 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 15 }, "clocks_sm_mhz": { "min": 1950.0, "median": 1965.0, "max": 1995.0, "mean": 1970.0, "n": 15 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 15 }, "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": 15 }, "busy_samples": 15 } } }, "energy_j_per_1k_tokens": 7295.49, "power_mean_w_all_cards": 418.76, "energy_j_per_1k_tokens_all_cards": 7295.49, "watts_per_tok_s": 7.2955, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 1950.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 418.76, "power_pct_of_limit": 93.1, "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": false, "verdict": "no clock drop during decode (draw held at 93.1% of the 450 W cap) (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 15, "temp_max_c": 56.0, "temp_min_c": 54.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 1950.0, "clock_max_mhz": 1995.0, "clock_median_mhz": 1965.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 450.0, "power_mean_w": 418.76, "power_pct_of_limit": 93.1, "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": false, "verdict": "no clock drop during decode (draw held at 93.1% of the 450 W cap) (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "mistral-small3.2:24b", "model_density": "dense", "weight_bytes": 15177384862, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": 871.2, "fraction_of_peak_pct": 88.5, "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 } }, "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": 57.368, "median": 57.396, "max": 57.4, "mean": 57.388, "n": 3 }, "ttft_ms": { "min": 271.37, "median": 275.63, "max": 284.3, "mean": 277.1, "n": 3 }, "power_mean_w": { "min": 408.83, "median": 418.76, "max": 426.1, "mean": 417.9, "n": 3 }, "power_mean_w_all_cards": { "min": 408.83, "median": 418.76, "max": 426.1, "mean": 417.9, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 7126.41, "median": 7295.49, "max": 7423.83, "mean": 7281.91, "n": 3 }, "energy_j_per_1k_tokens": { "min": 7126.41, "median": 7295.49, "max": 7423.83, "mean": 7281.91, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": { "min": 870.7, "median": 871.1, "max": 871.2, "mean": 871.0, "n": 3 }, "decode_bandwidth_fraction_of_peak_pct": { "min": 88.5, "median": 88.5, "max": 88.5, "mean": 88.5, "n": 3 }, "sw_power_cap_active_fraction": { "min": 0.875, "median": 1.0, "max": 1.0, "mean": 0.958, "n": 3 }, "temp_max_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.7, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 47.0, "max": 66.0, "mean": 37.7, "n": 3 } }, "finished_utc": "2026-09-17T23:38:06Z" }, { "num_ctx": 98304, "started_utc": "2026-09-17T23:38:06Z", "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 2550.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 23152.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20602.0 }, "runner_peak_rss_gib": 14.16, "runner_rss_within_allowance": false, "runner_procs": [ { "pid": 180583, "rss_kb": 14843248, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 2805 389 139 44532 44235\nSwap: 24575 795 23780", "kv_footprint_bytes": 10093904897, "kv_footprint_gib": 9.4, "finished_utc": "2026-09-17T23:38:12Z" } ], "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-1809d1e1-1372-f6aa-b687-77ac70675fc4": 23152.0 } }, "ups_after_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-17T23:38:12Z", "available": false, "error": "upsc rc=1: Error: Connection failure: Connection refused" }, "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": "2026-09-17T23:38:12Z" }