{ "model_id": "m4", "model_tag": "mistral-small3.2:24b", "arm": "one", "box_class": "desktop-3090", "mode": "context-ladder", "arm_cards": [ "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08" ], "target_uuid": "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08", "started_utc": "2026-09-18T00:38:30Z", "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 x16 slot of the same board 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 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 two GeForce RTX 3080 10 GB cards were measured the night before. A temperature here is a reading of THAT arrangement. It differs from the pair's arrangement in the way that matters most to a temperature: with one card there is no second board warming the air or blocking a face, so this card has open air on both sides where the pair did not. The card is traced on every run.", "card_labels": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": "index 0 - EVGA - GPU-7aa0be10" }, "card_identity": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": 0, "bus_id": "00000000:01:00.0", "pci_sub_device_id": "0x39753842", "vendor": "EVGA", "board_per_operator": "EVGA GeForce RTX 3090 XC3 Ultra 24 GB", "vbios": "94.02.42.80.C1", "serial": null, "pcie_link_width_max": 16, "pcie_link_width_negotiated": 16, "memory_total_mib": 24576, "power_default_limit_w": 350.0, "power_min_limit_w": 100.0, "power_max_limit_w": 366.0, "power_limit_at_bench_start_w": 250.0, "note": "the single card seated in this box's x16 slot for the control arm. It came up at 280 W from a boot unit written for the box's former pair (/etc/desktop/gpu-power-caps.conf, 2026-08-27) whose other row names a card that is no longer present; the operator set 250 W by hand before this bench and the read-back is recorded per arm." } }, "memory_temp_support": { "checked_utc": "2026-09-18T00:38:30Z", "paths": { "nvidia-smi --query-gpu=temperature.memory": { "raw": "0, GPU-7aa0be10-974f-6430-52fe-09018a0e2e08, N/A", "readable": false }, "nvidia-smi -q -d TEMPERATURE": { "lines": [ "GPU Shutdown Temp : 98 C", "GPU Slowdown Temp : 95 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": "2026-09-18T00:35:46Z", "unit": "bench-3090-one", "shape": "one", "port": 11470, "base_url": "http://127.0.0.1:11470", "cuda_visible_devices": "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08", "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, 350.00 W;", "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 16;" }, "ups_before_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-18T00:38:30Z", "available": false, "error": "upsc rc=1: Error: Connection failure: Connection refused" }, "levels": [ { "num_ctx": 4096, "started_utc": "2026-09-18T00:38:31Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 15044.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 313.0 }, "runner_peak_rss_gib": 0.8, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 18752, "rss_kb": 842632, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3358 7952 50 36301 43682\nSwap: 24575 0 24575", "kv_footprint_bytes": 0, "kv_footprint_gib": 0.0, "runs": [ { "ttft_ms": 286.81, "first_token_was_thinking": false, "wall_s": 5.725, "eval_count": 256, "eval_duration_ns": 5436296000, "decode_tok_s": 47.091, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 37314000, "prefill_tok_s": 27415.983, "load_duration_ns": 245576908, "total_duration_ns": 5723013734, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.891, "decode_tok_s_wall_gross": 44.716, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 129.85, "median": 129.89, "max": 129.99, "mean": 129.91, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 2.0, "mean": 0.3, "n": 6 }, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 129.91 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "power_w_all_cards": 129.91, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 284.4, "max_w": 311.38, "median_w": 298.79, "mean_w_whole_window": 269.68, "max_w_whole_window": 311.38, "limit_w": 350.0, "over_idle_w": 154.49, "samples": 21, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 129.77, "median": 298.71, "max": 311.38, "mean": 269.68, "n": 21 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.3, "n": 21 }, "temperature_c": { "min": 55.0, "median": 60.0, "max": 61.0, "mean": 59.3, "n": 21 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 21 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 }, "clocks_sm_mhz": { "min": 960.0, "median": 1260.0, "max": 1725.0, "mean": 1238.6, "n": 21 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 160.37, "median": 298.79, "max": 311.38, "mean": 284.4, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 19 }, "utilization_pct": { "min": 96.0, "median": 99.0, "max": 99.0, "mean": 98.7, "n": 19 }, "temperature_c": { "min": 59.0, "median": 60.0, "max": 61.0, "mean": 59.8, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 960.0, "median": 1185.0, "max": 1380.0, "mean": 1187.4, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6039.39, "power_mean_w_all_cards": 284.4, "energy_j_per_1k_tokens_all_cards": 6039.39, "watts_per_tok_s": 6.0394, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 61.0, "temp_min_c": 59.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.4, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.3% of the cap and the card at 61 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 61.0, "temp_min_c": 59.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.4, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.3% of the cap and the card at 61 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 317.11, "first_token_was_thinking": false, "wall_s": 5.757, "eval_count": 256, "eval_duration_ns": 5437971000, "decode_tok_s": 47.076, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 56957000, "prefill_tok_s": 17960.918, "load_duration_ns": 256288077, "total_duration_ns": 5754863434, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.876, "decode_tok_s_wall_gross": 44.468, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 23.09, "median": 23.24, "max": 23.44, "mean": 23.26, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.26 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "power_w_all_cards": 23.26, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 282.8, "max_w": 298.74, "median_w": 297.54, "mean_w_whole_window": 249.5, "max_w_whole_window": 298.74, "limit_w": 350.0, "over_idle_w": 259.54, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 22.73, "median": 297.43, "max": 298.74, "mean": 249.5, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 50.0, "median": 58.5, "max": 60.0, "mean": 57.6, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 43.0, "mean": 10.9, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1215.0, "max": 1350.0, "mean": 1151.6, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 94.89, "median": 297.54, "max": 298.74, "mean": 282.8, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 57.0, "median": 59.0, "max": 60.0, "mean": 58.5, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 39.0, "mean": 6.1, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 975.0, "median": 1230.0, "max": 1350.0, "mean": 1193.7, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6007.26, "power_mean_w_all_cards": 282.8, "energy_j_per_1k_tokens_all_cards": 6007.26, "watts_per_tok_s": 6.0073, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 60.0, "temp_min_c": 57.0, "temp_rise_c": 3.0, "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": 6.1, "fan_max_pct": 39.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1230.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 282.8, "power_pct_of_limit": 80.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 80.8% of the cap and the card at 60 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 60.0, "temp_min_c": 57.0, "temp_rise_c": 3.0, "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": 6.1, "fan_max_pct": 39.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1230.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 282.8, "power_pct_of_limit": 80.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 80.8% of the cap and the card at 60 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 270.29, "first_token_was_thinking": false, "wall_s": 5.7611, "eval_count": 256, "eval_duration_ns": 5489112000, "decode_tok_s": 46.638, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21842000, "prefill_tok_s": 46836.37, "load_duration_ns": 244400982, "total_duration_ns": 5759117607, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.441, "decode_tok_s_wall_gross": 44.436, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 128.83, "median": 128.86, "max": 128.91, "mean": 128.87, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 128.87 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "power_w_all_cards": 128.87, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 296.47, "max_w": 310.5, "median_w": 297.99, "mean_w_whole_window": 269.9, "max_w_whole_window": 310.5, "limit_w": 350.0, "over_idle_w": 167.6, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 128.85, "median": 297.64, "max": 310.5, "mean": 269.9, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 81.0, "n": 22 }, "temperature_c": { "min": 56.0, "median": 61.0, "max": 62.0, "mean": 60.8, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 960.0, "median": 1200.0, "max": 1725.0, "mean": 1210.9, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 261.21, "median": 297.99, "max": 310.5, "mean": 296.47, "n": 18 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15357.0, "median": 15357.0, "max": 15357.0, "mean": 15357.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 60.0, "median": 61.5, "max": 62.0, "mean": 61.4, "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": 960.0, "median": 1125.0, "max": 1320.0, "mean": 1142.5, "n": 18 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 6356.86, "power_mean_w_all_cards": 296.47, "energy_j_per_1k_tokens_all_cards": 6356.86, "watts_per_tok_s": 6.3568, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 60.0, "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": 960.0, "clock_max_mhz": 1320.0, "clock_median_mhz": 1125.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 296.47, "power_pct_of_limit": 84.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 84.7% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 60.0, "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": 960.0, "clock_max_mhz": 1320.0, "clock_median_mhz": 1125.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 296.47, "power_pct_of_limit": 84.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 84.7% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it (the fans never spun up -- zero-RPM idle mode)" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.638, "median": 47.076, "max": 47.091, "mean": 46.935, "n": 3 }, "ttft_ms": { "min": 270.29, "median": 286.81, "max": 317.11, "mean": 291.4, "n": 3 }, "power_mean_w": { "min": 282.8, "median": 284.4, "max": 296.47, "mean": 287.89, "n": 3 }, "power_mean_w_all_cards": { "min": 282.8, "median": 284.4, "max": 296.47, "mean": 287.89, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 6007.26, "median": 6039.39, "max": 6356.86, "mean": 6134.5, "n": 3 }, "energy_j_per_1k_tokens": { "min": 6007.26, "median": 6039.39, "max": 6356.86, "mean": 6134.5, "n": 3 } }, "finished_utc": "2026-09-18T00:39:48Z" }, { "num_ctx": 8192, "started_utc": "2026-09-18T00:39:48Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 588.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15357.0 }, "runner_peak_rss_gib": 0.84, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 19780, "rss_kb": 877540, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3386 7925 82 36334 43654\nSwap: 24575 0 24575", "kv_footprint_bytes": 599795958, "kv_footprint_gib": 0.56, "runs": [ { "ttft_ms": 274.0, "first_token_was_thinking": false, "wall_s": 5.7792, "eval_count": 256, "eval_duration_ns": 5503249000, "decode_tok_s": 46.518, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 28633000, "prefill_tok_s": 35728.006, "load_duration_ns": 242364118, "total_duration_ns": 5777063442, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.32, "decode_tok_s_wall_gross": 44.297, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 131.08, "median": 131.09, "max": 131.13, "mean": 131.1, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 131.1 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "power_w_all_cards": 131.1, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 20, "mean_w": 285.67, "max_w": 312.18, "median_w": 299.35, "mean_w_whole_window": 271.62, "max_w_whole_window": 312.18, "limit_w": 350.0, "over_idle_w": 154.57, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 131.13, "median": 299.13, "max": 312.18, "mean": 271.62, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 87.3, "n": 22 }, "temperature_c": { "min": 59.0, "median": 64.0, "max": 65.0, "mean": 63.6, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 54.0, "mean": 15.4, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 960.0, "median": 1177.5, "max": 1725.0, "mean": 1210.9, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 162.09, "median": 299.35, "max": 312.18, "mean": 285.67, "n": 20 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 20 }, "utilization_pct": { "min": 79.0, "median": 99.0, "max": 99.0, "mean": 96.0, "n": 20 }, "temperature_c": { "min": 63.0, "median": 64.0, "max": 65.0, "mean": 64.1, "n": 20 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 54.0, "mean": 16.9, "n": 20 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 20 }, "clocks_sm_mhz": { "min": 960.0, "median": 1155.0, "max": 1380.0, "mean": 1159.5, "n": 20 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 6141.07, "power_mean_w_all_cards": 285.67, "energy_j_per_1k_tokens_all_cards": 6141.07, "watts_per_tok_s": 6.1411, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 65.0, "temp_min_c": 63.0, "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": 16.9, "fan_max_pct": 54.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 285.67, "power_pct_of_limit": 81.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.6% of the cap and the card at 65 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 65.0, "temp_min_c": 63.0, "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": 16.9, "fan_max_pct": 54.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 285.67, "power_pct_of_limit": 81.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.6% of the cap and the card at 65 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 324.61, "first_token_was_thinking": false, "wall_s": 5.8072, "eval_count": 256, "eval_duration_ns": 5480790000, "decode_tok_s": 46.709, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 76416000, "prefill_tok_s": 13387.249, "load_duration_ns": 245213828, "total_duration_ns": 5805216270, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.511, "decode_tok_s_wall_gross": 44.083, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 23.05, "median": 23.13, "max": 23.75, "mean": 23.31, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.31 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "power_w_all_cards": 23.31, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 294.64, "max_w": 300.31, "median_w": 299.68, "mean_w_whole_window": 250.35, "max_w_whole_window": 300.31, "limit_w": 350.0, "over_idle_w": 271.33, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 23.33, "median": 299.25, "max": 300.31, "mean": 250.35, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 81.0, "n": 22 }, "temperature_c": { "min": 53.0, "median": 61.0, "max": 62.0, "mean": 60.5, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1177.5, "max": 1335.0, "mean": 1093.0, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 213.15, "median": 299.68, "max": 300.31, "mean": 294.64, "n": 18 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 60.0, "median": 62.0, "max": 62.0, "mean": 61.5, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 960.0, "median": 1177.5, "max": 1335.0, "mean": 1162.5, "n": 18 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 6308.05, "power_mean_w_all_cards": 294.64, "energy_j_per_1k_tokens_all_cards": 6308.05, "watts_per_tok_s": 6.308, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 60.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1177.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 294.64, "power_pct_of_limit": 84.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 84.2% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 60.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1177.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 294.64, "power_pct_of_limit": 84.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 84.2% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 274.7, "first_token_was_thinking": false, "wall_s": 5.7498, "eval_count": 256, "eval_duration_ns": 5473033000, "decode_tok_s": 46.775, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21756000, "prefill_tok_s": 47021.511, "load_duration_ns": 249713594, "total_duration_ns": 5747495534, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.575, "decode_tok_s_wall_gross": 44.523, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 130.24, "median": 130.33, "max": 130.39, "mean": 130.32, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 130.32 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "power_w_all_cards": 130.32, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 292.93, "max_w": 313.03, "median_w": 299.78, "mean_w_whole_window": 272.4, "max_w_whole_window": 313.03, "limit_w": 350.0, "over_idle_w": 162.61, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 130.13, "median": 299.12, "max": 313.03, "mean": 272.4, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 57.0, "median": 62.0, "max": 62.0, "mean": 61.2, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 945.0, "median": 1200.0, "max": 1725.0, "mean": 1218.4, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 187.43, "median": 299.78, "max": 313.03, "mean": 292.93, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 15945.0, "median": 15945.0, "max": 15945.0, "mean": 15945.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 61.0, "median": 62.0, "max": 62.0, "mean": 61.7, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 945.0, "median": 1155.0, "max": 1365.0, "mean": 1160.5, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6262.56, "power_mean_w_all_cards": 292.93, "energy_j_per_1k_tokens_all_cards": 6262.56, "watts_per_tok_s": 6.2625, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 62.0, "temp_min_c": 61.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1365.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 292.93, "power_pct_of_limit": 83.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.7% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 62.0, "temp_min_c": 61.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1365.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 292.93, "power_pct_of_limit": 83.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.7% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.518, "median": 46.709, "max": 46.775, "mean": 46.667, "n": 3 }, "ttft_ms": { "min": 274.0, "median": 274.7, "max": 324.61, "mean": 291.1, "n": 3 }, "power_mean_w": { "min": 285.67, "median": 292.93, "max": 294.64, "mean": 291.08, "n": 3 }, "power_mean_w_all_cards": { "min": 285.67, "median": 292.93, "max": 294.64, "mean": 291.08, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 6141.07, "median": 6262.56, "max": 6308.05, "mean": 6237.23, "n": 3 }, "energy_j_per_1k_tokens": { "min": 6141.07, "median": 6262.56, "max": 6308.05, "mean": 6237.23, "n": 3 } }, "finished_utc": "2026-09-18T00:40:58Z" }, { "num_ctx": 16384, "started_utc": "2026-09-18T00:40:58Z", "loads": true, "size_total": 15932598517, "size_vram": 15932598517, "context_length_loaded": 16384, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 696.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 15945.0 }, "runner_peak_rss_gib": 0.85, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 20667, "rss_kb": 894224, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3353 7957 98 36350 43687\nSwap: 24575 0 24575", "kv_footprint_bytes": 1329604854, "kv_footprint_gib": 1.24, "runs": [ { "ttft_ms": 327.76, "first_token_was_thinking": false, "wall_s": 5.8095, "eval_count": 256, "eval_duration_ns": 5479999000, "decode_tok_s": 46.715, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 81296000, "prefill_tok_s": 12583.645, "load_duration_ns": 243518432, "total_duration_ns": 5807577608, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.518, "decode_tok_s_wall_gross": 44.066, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 23.02, "median": 23.1, "max": 23.82, "mean": 23.31, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.31 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "power_w_all_cards": 23.31, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 293.26, "max_w": 300.12, "median_w": 298.06, "mean_w_whole_window": 248.98, "max_w_whole_window": 300.12, "limit_w": 350.0, "over_idle_w": 269.95, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 23.1, "median": 297.84, "max": 300.12, "mean": 248.98, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 81.0, "n": 22 }, "temperature_c": { "min": 51.0, "median": 60.0, "max": 61.0, "mean": 59.0, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 40.0, "median": 42.0, "max": 47.0, "mean": 42.6, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1185.0, "max": 1350.0, "mean": 1096.4, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 209.64, "median": 298.06, "max": 300.12, "mean": 293.26, "n": 18 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 18 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 18 }, "temperature_c": { "min": 59.0, "median": 60.0, "max": 61.0, "mean": 60.1, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 40.0, "median": 41.0, "max": 45.0, "mean": 41.7, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 960.0, "median": 1185.0, "max": 1350.0, "mean": 1165.0, "n": 18 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 6277.6, "power_mean_w_all_cards": 293.26, "energy_j_per_1k_tokens_all_cards": 6277.6, "watts_per_tok_s": 6.2776, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 61.0, "temp_min_c": 59.0, "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": 41.7, "fan_max_pct": 45.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.26, "power_pct_of_limit": 83.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.8% of the cap and the card at 61 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 61.0, "temp_min_c": 59.0, "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": 41.7, "fan_max_pct": 45.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.26, "power_pct_of_limit": 83.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.8% of the cap and the card at 61 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 273.6, "first_token_was_thinking": false, "wall_s": 5.7826, "eval_count": 256, "eval_duration_ns": 5507283000, "decode_tok_s": 46.484, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21855000, "prefill_tok_s": 46808.511, "load_duration_ns": 248870470, "total_duration_ns": 5780679756, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.288, "decode_tok_s_wall_gross": 44.271, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 130.29, "median": 130.33, "max": 130.35, "mean": 130.32, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 130.32 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "power_w_all_cards": 130.32, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 291.46, "max_w": 308.07, "median_w": 298.51, "mean_w_whole_window": 270.96, "max_w_whole_window": 308.07, "limit_w": 350.0, "over_idle_w": 161.14, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 130.34, "median": 298.18, "max": 308.07, "mean": 270.96, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 57.0, "median": 62.0, "max": 63.0, "mean": 61.7, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 35.0, "median": 36.0, "max": 36.0, "mean": 35.5, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 960.0, "median": 1170.0, "max": 1725.0, "mean": 1208.9, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 185.57, "median": 298.51, "max": 308.07, "mean": 291.46, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 61.0, "median": 62.0, "max": 63.0, "mean": 62.2, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 35.0, "median": 35.0, "max": 36.0, "mean": 35.5, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 960.0, "median": 1155.0, "max": 1335.0, "mean": 1147.1, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6270.13, "power_mean_w_all_cards": 291.46, "energy_j_per_1k_tokens_all_cards": 6270.13, "watts_per_tok_s": 6.2701, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 63.0, "temp_min_c": 61.0, "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": 35.5, "fan_max_pct": 36.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 291.46, "power_pct_of_limit": 83.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.3% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 63.0, "temp_min_c": 61.0, "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": 35.5, "fan_max_pct": 36.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 291.46, "power_pct_of_limit": 83.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.3% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 308.73, "first_token_was_thinking": false, "wall_s": 5.8416, "eval_count": 256, "eval_duration_ns": 5531010000, "decode_tok_s": 46.284, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 56707000, "prefill_tok_s": 18040.101, "load_duration_ns": 249146469, "total_duration_ns": 5839619515, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.088, "decode_tok_s_wall_gross": 43.824, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.91, "median": 23.05, "max": 24.45, "mean": 23.47, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.47 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "power_w_all_cards": 23.47, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 20, "mean_w": 273.24, "max_w": 300.17, "median_w": 298.96, "mean_w_whole_window": 250.59, "max_w_whole_window": 300.17, "limit_w": 350.0, "over_idle_w": 249.77, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 22.97, "median": 298.79, "max": 300.17, "mean": 250.59, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.5, "n": 22 }, "temperature_c": { "min": 55.0, "median": 63.0, "max": 64.0, "mean": 62.2, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 33.0, "median": 34.0, "max": 34.0, "mean": 33.5, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1140.0, "max": 1335.0, "mean": 1088.9, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 57.58, "median": 298.96, "max": 300.17, "mean": 273.24, "n": 20 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 16641.0, "median": 16641.0, "max": 16641.0, "mean": 16641.0, "n": 20 }, "utilization_pct": { "min": 94.0, "median": 99.0, "max": 99.0, "mean": 98.5, "n": 20 }, "temperature_c": { "min": 59.0, "median": 63.0, "max": 64.0, "mean": 62.9, "n": 20 }, "memory_temperature_c": null, "fan_pct": { "min": 33.0, "median": 33.5, "max": 34.0, "mean": 33.5, "n": 20 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 20 }, "clocks_sm_mhz": { "min": 945.0, "median": 1177.5, "max": 1335.0, "mean": 1162.5, "n": 20 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 5903.49, "power_mean_w_all_cards": 273.24, "energy_j_per_1k_tokens_all_cards": 5903.49, "watts_per_tok_s": 5.9036, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 64.0, "temp_min_c": 59.0, "temp_rise_c": 5.0, "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": 33.5, "fan_max_pct": 34.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1177.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 273.24, "power_pct_of_limit": 78.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 78.1% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 64.0, "temp_min_c": 59.0, "temp_rise_c": 5.0, "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": 33.5, "fan_max_pct": 34.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1177.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 273.24, "power_pct_of_limit": 78.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 78.1% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.284, "median": 46.484, "max": 46.715, "mean": 46.494, "n": 3 }, "ttft_ms": { "min": 273.6, "median": 308.73, "max": 327.76, "mean": 303.36, "n": 3 }, "power_mean_w": { "min": 273.24, "median": 291.46, "max": 293.26, "mean": 285.99, "n": 3 }, "power_mean_w_all_cards": { "min": 273.24, "median": 291.46, "max": 293.26, "mean": 285.99, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 5903.49, "median": 6270.13, "max": 6277.6, "mean": 6150.41, "n": 3 }, "energy_j_per_1k_tokens": { "min": 5903.49, "median": 6270.13, "max": 6277.6, "mean": 6150.41, "n": 3 } }, "finished_utc": "2026-09-18T00:42:19Z" }, { "num_ctx": 32768, "started_utc": "2026-09-18T00:42:19Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 1124.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 16641.0 }, "runner_peak_rss_gib": 0.83, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 21748, "rss_kb": 872396, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3376 7933 78 36331 43664\nSwap: 24575 0 24575", "kv_footprint_bytes": 2524971008, "kv_footprint_gib": 2.35, "runs": [ { "ttft_ms": 277.3, "first_token_was_thinking": false, "wall_s": 5.7684, "eval_count": 256, "eval_duration_ns": 5489536000, "decode_tok_s": 46.634, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 29417000, "prefill_tok_s": 34775.81, "load_duration_ns": 244930492, "total_duration_ns": 5766531397, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.438, "decode_tok_s_wall_gross": 44.379, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 131.61, "median": 131.72, "max": 131.8, "mean": 131.71, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 131.71 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "power_w_all_cards": 131.71, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 293.4, "max_w": 312.3, "median_w": 300.45, "mean_w_whole_window": 273.13, "max_w_whole_window": 312.3, "limit_w": 350.0, "over_idle_w": 161.69, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 131.66, "median": 300.11, "max": 312.3, "mean": 273.13, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 59.0, "median": 64.0, "max": 64.0, "mean": 63.2, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 51.0, "max": 51.0, "mean": 51.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 960.0, "median": 1200.0, "max": 1725.0, "mean": 1216.4, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 186.71, "median": 300.45, "max": 312.3, "mean": 293.4, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 63.0, "median": 64.0, "max": 64.0, "mean": 63.6, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 51.0, "max": 51.0, "mean": 51.0, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 960.0, "median": 1140.0, "max": 1335.0, "mean": 1154.2, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6291.52, "power_mean_w_all_cards": 293.4, "energy_j_per_1k_tokens_all_cards": 6291.52, "watts_per_tok_s": 6.2915, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 64.0, "temp_min_c": 63.0, "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": 51.0, "fan_max_pct": 51.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1140.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.4, "power_pct_of_limit": 83.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.8% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 64.0, "temp_min_c": 63.0, "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": 51.0, "fan_max_pct": 51.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1140.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.4, "power_pct_of_limit": 83.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.8% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 270.37, "first_token_was_thinking": false, "wall_s": 5.763, "eval_count": 256, "eval_duration_ns": 5490688000, "decode_tok_s": 46.624, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21667000, "prefill_tok_s": 47214.658, "load_duration_ns": 245517290, "total_duration_ns": 5760709568, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.426, "decode_tok_s_wall_gross": 44.421, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 131.61, "median": 131.65, "max": 131.69, "mean": 131.65, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 131.65 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "power_w_all_cards": 131.65, "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": 20, "mean_w": 287.48, "max_w": 313.52, "median_w": 300.56, "mean_w_whole_window": 273.56, "max_w_whole_window": 313.52, "limit_w": 350.0, "over_idle_w": 155.83, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 131.61, "median": 300.24, "max": 313.52, "mean": 273.56, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.7, "n": 22 }, "temperature_c": { "min": 58.0, "median": 63.0, "max": 64.0, "mean": 62.8, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 52.0, "max": 52.0, "mean": 51.7, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 960.0, "median": 1177.5, "max": 1725.0, "mean": 1210.9, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 164.13, "median": 300.56, "max": 313.52, "mean": 287.48, "n": 20 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 20 }, "utilization_pct": { "min": 96.0, "median": 99.0, "max": 99.0, "mean": 98.7, "n": 20 }, "temperature_c": { "min": 62.0, "median": 63.0, "max": 64.0, "mean": 63.2, "n": 20 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 52.0, "max": 52.0, "mean": 51.6, "n": 20 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 20 }, "clocks_sm_mhz": { "min": 960.0, "median": 1155.0, "max": 1380.0, "mean": 1163.2, "n": 20 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 6165.87, "power_mean_w_all_cards": 287.48, "energy_j_per_1k_tokens_all_cards": 6165.87, "watts_per_tok_s": 6.1659, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 64.0, "temp_min_c": 62.0, "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": 51.6, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 287.48, "power_pct_of_limit": 82.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 82.1% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 64.0, "temp_min_c": 62.0, "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": 51.6, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 287.48, "power_pct_of_limit": 82.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 82.1% of the cap and the card at 64 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 319.49, "first_token_was_thinking": false, "wall_s": 5.7773, "eval_count": 256, "eval_duration_ns": 5455947000, "decode_tok_s": 46.921, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 75194000, "prefill_tok_s": 13604.809, "load_duration_ns": 241339453, "total_duration_ns": 5775287262, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.722, "decode_tok_s_wall_gross": 44.311, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 23.44, "median": 23.61, "max": 23.83, "mean": 23.63, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.63 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "power_w_all_cards": 23.63, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 284.53, "max_w": 300.26, "median_w": 299.26, "mean_w_whole_window": 250.65, "max_w_whole_window": 300.26, "limit_w": 350.0, "over_idle_w": 260.9, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 23.14, "median": 299.1, "max": 300.26, "mean": 250.65, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.2, "n": 22 }, "temperature_c": { "min": 53.0, "median": 61.0, "max": 62.0, "mean": 60.0, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 51.0, "mean": 49.4, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1177.5, "max": 1335.0, "mean": 1116.1, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 100.22, "median": 299.26, "max": 300.26, "mean": 284.53, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 17765.0, "median": 17765.0, "max": 17765.0, "mean": 17765.0, "n": 19 }, "utilization_pct": { "min": 98.0, "median": 99.0, "max": 99.0, "mean": 98.7, "n": 19 }, "temperature_c": { "min": 59.0, "median": 61.0, "max": 62.0, "mean": 60.8, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 51.0, "mean": 49.1, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 975.0, "median": 1185.0, "max": 1335.0, "mean": 1186.6, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6063.99, "power_mean_w_all_cards": 284.53, "energy_j_per_1k_tokens_all_cards": 6063.99, "watts_per_tok_s": 6.064, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 62.0, "temp_min_c": 59.0, "temp_rise_c": 3.0, "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": 49.1, "fan_max_pct": 51.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.53, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 81.3% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 62.0, "temp_min_c": 59.0, "temp_rise_c": 3.0, "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": 49.1, "fan_max_pct": 51.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1185.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.53, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 81.3% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.624, "median": 46.634, "max": 46.921, "mean": 46.726, "n": 3 }, "ttft_ms": { "min": 270.37, "median": 277.3, "max": 319.49, "mean": 289.05, "n": 3 }, "power_mean_w": { "min": 284.53, "median": 287.48, "max": 293.4, "mean": 288.47, "n": 3 }, "power_mean_w_all_cards": { "min": 284.53, "median": 287.48, "max": 293.4, "mean": 288.47, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 6063.99, "median": 6165.87, "max": 6291.52, "mean": 6173.79, "n": 3 }, "energy_j_per_1k_tokens": { "min": 6063.99, "median": 6165.87, "max": 6291.52, "mean": 6173.79, "n": 3 } }, "finished_utc": "2026-09-18T00:43:30Z" }, { "num_ctx": 49152, "started_utc": "2026-09-18T00:43:30Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 1384.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 17765.0 }, "runner_peak_rss_gib": 0.85, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 22773, "rss_kb": 889636, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3276 8033 94 36347 43763\nSwap: 24575 0 24575", "kv_footprint_bytes": 3976200192, "kv_footprint_gib": 3.7, "runs": [ { "ttft_ms": 275.43, "first_token_was_thinking": false, "wall_s": 5.7772, "eval_count": 256, "eval_duration_ns": 5499742000, "decode_tok_s": 46.548, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 29163000, "prefill_tok_s": 35078.696, "load_duration_ns": 243334666, "total_duration_ns": 5775041482, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.349, "decode_tok_s_wall_gross": 44.313, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 130.55, "median": 130.66, "max": 130.71, "mean": 130.64, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 130.64 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "power_w_all_cards": 130.64, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 20, "mean_w": 285.2, "max_w": 311.32, "median_w": 299.19, "mean_w_whole_window": 271.21, "max_w_whole_window": 311.32, "limit_w": 350.0, "over_idle_w": 154.56, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 130.77, "median": 299.08, "max": 311.32, "mean": 271.21, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.5, "n": 22 }, "temperature_c": { "min": 57.0, "median": 62.0, "max": 63.0, "mean": 61.8, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 43.0, "median": 44.0, "max": 45.0, "mean": 43.7, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 945.0, "median": 1185.0, "max": 1725.0, "mean": 1210.9, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 158.91, "median": 299.19, "max": 311.32, "mean": 285.2, "n": 20 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 20 }, "utilization_pct": { "min": 93.0, "median": 99.0, "max": 99.0, "mean": 98.4, "n": 20 }, "temperature_c": { "min": 61.0, "median": 62.0, "max": 63.0, "mean": 62.3, "n": 20 }, "memory_temperature_c": null, "fan_pct": { "min": 43.0, "median": 44.0, "max": 44.0, "mean": 43.6, "n": 20 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 20 }, "clocks_sm_mhz": { "min": 945.0, "median": 1155.0, "max": 1380.0, "mean": 1159.5, "n": 20 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 6127.06, "power_mean_w_all_cards": 285.2, "energy_j_per_1k_tokens_all_cards": 6127.06, "watts_per_tok_s": 6.127, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 63.0, "temp_min_c": 61.0, "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": 43.6, "fan_max_pct": 44.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 285.2, "power_pct_of_limit": 81.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.5% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 63.0, "temp_min_c": 61.0, "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": 43.6, "fan_max_pct": 44.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1380.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 285.2, "power_pct_of_limit": 81.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 81.5% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 319.31, "first_token_was_thinking": false, "wall_s": 5.8358, "eval_count": 256, "eval_duration_ns": 5514618000, "decode_tok_s": 46.422, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 73217000, "prefill_tok_s": 13972.165, "load_duration_ns": 243106042, "total_duration_ns": 5833631490, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.225, "decode_tok_s_wall_gross": 43.867, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.8, "median": 22.82, "max": 23.31, "mean": 22.98, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 22.98 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "power_w_all_cards": 22.98, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 284.21, "max_w": 299.68, "median_w": 298.77, "mean_w_whole_window": 250.19, "max_w_whole_window": 299.68, "limit_w": 350.0, "over_idle_w": 261.23, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 23.13, "median": 298.12, "max": 299.68, "mean": 250.19, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 54.0, "median": 62.0, "max": 63.0, "mean": 61.3, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 37.0, "median": 38.0, "max": 40.0, "mean": 38.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1147.5, "max": 1335.0, "mean": 1081.4, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 98.47, "median": 298.77, "max": 299.68, "mean": 284.21, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 60.0, "median": 62.0, "max": 63.0, "mean": 62.2, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 37.0, "median": 38.0, "max": 39.0, "mean": 37.7, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 960.0, "median": 1155.0, "max": 1335.0, "mean": 1158.9, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6122.3, "power_mean_w_all_cards": 284.21, "energy_j_per_1k_tokens_all_cards": 6122.3, "watts_per_tok_s": 6.1223, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 63.0, "temp_min_c": 60.0, "temp_rise_c": 3.0, "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": 37.7, "fan_max_pct": 39.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.21, "power_pct_of_limit": 81.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 81.2% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 63.0, "temp_min_c": 60.0, "temp_rise_c": 3.0, "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": 37.7, "fan_max_pct": 39.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1155.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 284.21, "power_pct_of_limit": 81.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 81.2% of the cap and the card at 63 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 273.94, "first_token_was_thinking": false, "wall_s": 5.7816, "eval_count": 256, "eval_duration_ns": 5505915000, "decode_tok_s": 46.495, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21838000, "prefill_tok_s": 46844.949, "load_duration_ns": 248991821, "total_duration_ns": 5779550308, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.299, "decode_tok_s_wall_gross": 44.278, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 131.39, "median": 131.42, "max": 131.43, "mean": 131.41, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 131.41 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "power_w_all_cards": 131.41, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 293.09, "max_w": 311.64, "median_w": 300.37, "mean_w_whole_window": 272.68, "max_w_whole_window": 311.64, "limit_w": 350.0, "over_idle_w": 161.68, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 131.4, "median": 300.01, "max": 311.64, "mean": 272.68, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 85.5, "n": 22 }, "temperature_c": { "min": 59.0, "median": 64.0, "max": 65.0, "mean": 63.6, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 38.0, "median": 38.0, "max": 46.0, "mean": 39.6, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 945.0, "median": 1177.5, "max": 1725.0, "mean": 1212.3, "n": 22 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 22 }, "throttle_sw_power_cap_fraction": 0.8636, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 189.5, "median": 300.37, "max": 311.64, "mean": 293.09, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19149.0, "median": 19149.0, "max": 19149.0, "mean": 19149.0, "n": 19 }, "utilization_pct": { "min": 99.0, "median": 99.0, "max": 99.0, "mean": 99.0, "n": 19 }, "temperature_c": { "min": 63.0, "median": 64.0, "max": 65.0, "mean": 64.2, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 38.0, "median": 38.0, "max": 46.0, "mean": 39.9, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 945.0, "median": 1140.0, "max": 1320.0, "mean": 1151.1, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 6303.63, "power_mean_w_all_cards": 293.09, "energy_j_per_1k_tokens_all_cards": 6303.63, "watts_per_tok_s": 6.3037, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 65.0, "temp_min_c": 63.0, "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": 39.9, "fan_max_pct": 46.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1320.0, "clock_median_mhz": 1140.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.09, "power_pct_of_limit": 83.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.7% of the cap and the card at 65 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 65.0, "temp_min_c": 63.0, "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": 39.9, "fan_max_pct": 46.0, "clock_floor_mhz": 945.0, "clock_max_mhz": 1320.0, "clock_median_mhz": 1140.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 293.09, "power_pct_of_limit": 83.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.7% of the cap and the card at 65 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.422, "median": 46.495, "max": 46.548, "mean": 46.488, "n": 3 }, "ttft_ms": { "min": 273.94, "median": 275.43, "max": 319.31, "mean": 289.56, "n": 3 }, "power_mean_w": { "min": 284.21, "median": 285.2, "max": 293.09, "mean": 287.5, "n": 3 }, "power_mean_w_all_cards": { "min": 284.21, "median": 285.2, "max": 293.09, "mean": 287.5, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 6122.3, "median": 6127.06, "max": 6303.63, "mean": 6184.33, "n": 3 }, "energy_j_per_1k_tokens": { "min": 6122.3, "median": 6127.06, "max": 6303.63, "mean": 6184.33, "n": 3 } }, "finished_utc": "2026-09-18T00:44:41Z" }, { "num_ctx": 65536, "started_utc": "2026-09-18T00:44:41Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 1440.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 19149.0 }, "runner_peak_rss_gib": 0.86, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 23648, "rss_kb": 906728, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3370 7939 110 36363 43670\nSwap: 24575 0 24575", "kv_footprint_bytes": 5486149632, "kv_footprint_gib": 5.11, "runs": [ { "ttft_ms": 331.14, "first_token_was_thinking": false, "wall_s": 5.805, "eval_count": 256, "eval_duration_ns": 5472060000, "decode_tok_s": 46.783, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 81026000, "prefill_tok_s": 12625.577, "load_duration_ns": 247100564, "total_duration_ns": 5803062788, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.585, "decode_tok_s_wall_gross": 44.1, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 23.2, "median": 23.28, "max": 23.57, "mean": 23.35, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23.35 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "power_w_all_cards": 23.35, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 20, "mean_w": 273.35, "max_w": 300.8, "median_w": 299.59, "mean_w_whole_window": 250.66, "max_w_whole_window": 300.8, "limit_w": 350.0, "over_idle_w": 250.0, "samples": 22, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 23.1, "median": 299.55, "max": 300.8, "mean": 250.66, "n": 22 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 22 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 89.5, "n": 22 }, "temperature_c": { "min": 53.0, "median": 61.0, "max": 62.0, "mean": 60.5, "n": 22 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 22 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 }, "clocks_sm_mhz": { "min": 210.0, "median": 1177.5, "max": 1335.0, "mean": 1098.4, "n": 22 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9087.5, "n": 22 }, "throttle_sw_power_cap_fraction": 0.9091, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 22, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 62.63, "median": 299.59, "max": 300.8, "mean": 273.35, "n": 20 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 20 }, "utilization_pct": { "min": 94.0, "median": 99.0, "max": 99.0, "mean": 98.5, "n": 20 }, "temperature_c": { "min": 58.0, "median": 61.0, "max": 62.0, "mean": 61.1, "n": 20 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 20 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 20 }, "clocks_sm_mhz": { "min": 960.0, "median": 1215.0, "max": 1335.0, "mean": 1184.2, "n": 20 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 20 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 20 }, "busy_samples": 20 } } }, "energy_j_per_1k_tokens": 5842.92, "power_mean_w_all_cards": 273.35, "energy_j_per_1k_tokens_all_cards": 5842.92, "watts_per_tok_s": 5.8429, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 62.0, "temp_min_c": 58.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1215.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 273.35, "power_pct_of_limit": 78.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 78.1% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 20, "temp_max_c": 62.0, "temp_min_c": 58.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 960.0, "clock_max_mhz": 1335.0, "clock_median_mhz": 1215.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 273.35, "power_pct_of_limit": 78.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 78.1% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 22 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 1 }, { "ttft_ms": 268.05, "first_token_was_thinking": false, "wall_s": 5.7306, "eval_count": 256, "eval_duration_ns": 5460697000, "decode_tok_s": 46.88, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 21876000, "prefill_tok_s": 46763.577, "load_duration_ns": 243118567, "total_duration_ns": 5728527383, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.682, "decode_tok_s_wall_gross": 44.673, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 130.21, "median": 130.23, "max": 130.26, "mean": 130.23, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 130.23 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "power_w_all_cards": 130.23, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 18, "mean_w": 292.34, "max_w": 311.97, "median_w": 299.91, "mean_w_whole_window": 270.86, "max_w_whole_window": 311.97, "limit_w": 350.0, "over_idle_w": 162.11, "samples": 21, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 130.12, "median": 299.44, "max": 311.97, "mean": 270.86, "n": 21 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 84.8, "n": 21 }, "temperature_c": { "min": 57.0, "median": 61.0, "max": 62.0, "mean": 61.0, "n": 21 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 21 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 }, "clocks_sm_mhz": { "min": 975.0, "median": 1215.0, "max": 1725.0, "mean": 1222.1, "n": 21 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 21 }, "throttle_sw_power_cap_fraction": 0.8571, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 188.37, "median": 299.91, "max": 311.97, "mean": 292.34, "n": 18 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 18 }, "utilization_pct": { "min": 98.0, "median": 99.0, "max": 99.0, "mean": 98.9, "n": 18 }, "temperature_c": { "min": 61.0, "median": 62.0, "max": 62.0, "mean": 61.6, "n": 18 }, "memory_temperature_c": null, "fan_pct": { "min": 52.0, "median": 52.0, "max": 52.0, "mean": 52.0, "n": 18 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 18 }, "clocks_sm_mhz": { "min": 975.0, "median": 1132.5, "max": 1365.0, "mean": 1161.7, "n": 18 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 18 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 18 }, "busy_samples": 18 } } }, "energy_j_per_1k_tokens": 6235.86, "power_mean_w_all_cards": 292.34, "energy_j_per_1k_tokens_all_cards": 6235.86, "watts_per_tok_s": 6.2359, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 61.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1365.0, "clock_median_mhz": 1132.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 292.34, "power_pct_of_limit": 83.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.5% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 18, "temp_max_c": 62.0, "temp_min_c": 61.0, "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": 52.0, "fan_max_pct": 52.0, "clock_floor_mhz": 975.0, "clock_max_mhz": 1365.0, "clock_median_mhz": 1132.5, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 292.34, "power_pct_of_limit": 83.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "NEITHER cap: the SM clock varied with draw at 83.5% of the cap and the card at 62 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 2 }, { "ttft_ms": 294.27, "first_token_was_thinking": false, "wall_s": 5.731, "eval_count": 256, "eval_duration_ns": 5434646000, "decode_tok_s": 47.105, "prompt_eval_count": 1023, "prompt_eval_duration_ns": 49783000, "prefill_tok_s": 20549.183, "load_duration_ns": 241550343, "total_duration_ns": 5728844595, "streamed_chunks": 256, "response_chars": 1036, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 46.903, "decode_tok_s_wall_gross": 44.67, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 22.43, "median": 22.44, "max": 22.74, "mean": 22.54, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 6 }, "power_w_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 22.54 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "power_w_all_cards": 22.54, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 19, "mean_w": 271.7, "max_w": 298.41, "median_w": 297.22, "mean_w_whole_window": 248.08, "max_w_whole_window": 298.41, "limit_w": 350.0, "over_idle_w": 249.16, "samples": 21, "per_gpu": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 22.53, "median": 297.16, "max": 298.41, "mean": 248.08, "n": 21 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 21 }, "utilization_pct": { "min": 0.0, "median": 99.0, "max": 99.0, "mean": 88.4, "n": 21 }, "temperature_c": { "min": 50.0, "median": 58.0, "max": 60.0, "mean": 57.6, "n": 21 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 21 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 }, "clocks_sm_mhz": { "min": 210.0, "median": 1185.0, "max": 1350.0, "mean": 1136.4, "n": 21 }, "clocks_mem_mhz": { "min": 405.0, "median": 9501.0, "max": 9501.0, "mean": 9067.9, "n": 21 }, "throttle_sw_power_cap_fraction": 0.9048, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 21, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090", "power_w": { "min": 74.07, "median": 297.22, "max": 298.41, "mean": 271.7, "n": 19 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20589.0, "median": 20589.0, "max": 20589.0, "mean": 20589.0, "n": 19 }, "utilization_pct": { "min": 93.0, "median": 99.0, "max": 99.0, "mean": 97.7, "n": 19 }, "temperature_c": { "min": 55.0, "median": 59.0, "max": 60.0, "mean": 58.3, "n": 19 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 19 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 19 }, "clocks_sm_mhz": { "min": 975.0, "median": 1260.0, "max": 1350.0, "mean": 1200.0, "n": 19 }, "clocks_mem_mhz": { "min": 9501.0, "median": 9501.0, "max": 9501.0, "mean": 9501.0, "n": 19 }, "throttle_sw_power_cap_fraction": 1.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 19 }, "busy_samples": 19 } } }, "energy_j_per_1k_tokens": 5767.94, "power_mean_w_all_cards": 271.7, "energy_j_per_1k_tokens_all_cards": 5767.94, "watts_per_tok_s": 5.768, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 60.0, "temp_min_c": 55.0, "temp_rise_c": 5.0, "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": 975.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1260.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 271.7, "power_pct_of_limit": 77.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 77.6% of the cap and the card at 60 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 19, "temp_max_c": 60.0, "temp_min_c": 55.0, "temp_rise_c": 5.0, "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": 975.0, "clock_max_mhz": 1350.0, "clock_median_mhz": 1260.0, "mem_clock_floor_mhz": 9501.0, "mem_clock_max_mhz": 9501.0, "mem_clock_held": true, "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 271.7, "power_pct_of_limit": 77.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 1.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": true, "verdict": "NEITHER cap: the SM clock varied with draw at 77.6% of the cap and the card at 60 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it (the fans never spun up -- zero-RPM idle mode)" } }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "pcie_width_under_load": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 21 } }, "ups_during_run": { "ups": "pr1500@localhost", "samples": 0, "load_pct": null, "whole_box_realpower_w": null, "whole_box_note": "the UPS's reading of everything it feeds, not the cards; nvidia-smi's per-card watts are a different measurement of a smaller thing", "stop_threshold_pct": 80.0, "breached": false }, "stop_conditions_hit": [], "run": 3 } ], "derived": { "decode_tok_s": { "min": 46.783, "median": 46.88, "max": 47.105, "mean": 46.923, "n": 3 }, "ttft_ms": { "min": 268.05, "median": 294.27, "max": 331.14, "mean": 297.82, "n": 3 }, "power_mean_w": { "min": 271.7, "median": 273.35, "max": 292.34, "mean": 279.13, "n": 3 }, "power_mean_w_all_cards": { "min": 271.7, "median": 273.35, "max": 292.34, "mean": 279.13, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 5767.94, "median": 5842.92, "max": 6235.86, "mean": 5948.91, "n": 3 }, "energy_j_per_1k_tokens": { "min": 5767.94, "median": 5842.92, "max": 6235.86, "mean": 5948.91, "n": 3 } }, "finished_utc": "2026-09-18T00:46:12Z" }, { "num_ctx": 98304, "started_utc": "2026-09-18T00:46:12Z", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 2550.0 }, "memory_used_mib_per_card": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 23139.0 }, "memory_used_mib_per_card_before": { "GPU-7aa0be10-974f-6430-52fe-09018a0e2e08": 20589.0 }, "runner_peak_rss_gib": 14.16, "runner_rss_within_allowance": false, "runner_procs": [ { "pid": 24769, "rss_kb": 14843548, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3639 7669 142 36396 43401\nSwap: 24575 0 24575", "kv_footprint_bytes": 10093904897, "kv_footprint_gib": 9.4, "finished_utc": "2026-09-18T00:46:17Z" } ], "errors": [], "context_length_max": 131072, "quantization": "Q4_K_M", "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-7aa0be10-974f-6430-52fe-09018a0e2e08": 23139.0 } }, "ups_after_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-18T00:46:17Z", "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": "2026-09-18T00:46:17Z" }