{ "model_id": "m5", "model_tag": "gemma4:26b", "arm": "one", "box_class": "desktop-3090ti", "mode": "context-ladder", "arm_cards": [ "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4" ], "target_uuid": "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4", "started_utc": "2026-09-17T20:02:59Z", "prompt_sha256": "90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c", "num_predict": 256, "ladder_note": "The frozen prompt is about 537 tokens. Climbing this ladder measures the cost of RESERVING a context window, NOT the cost of filling one: a 128k window with 537 tokens in it decodes at nearly the speed of an 8k window with 537 tokens in it. What changes is the VRAM the reservation costs, and therefore what fits. Any tok/s difference across levels is a second-order effect of the allocation and is not a long-prompt number. The FILLED-WINDOW arm in this same directory is the one that answers the long-prompt question, and the two must never be quoted as if they were the same measurement.", "fit_rule_note": "A context rung FITS when `/api/ps` reports `size_vram == size` -- every byte of weights and KV on the card, nothing in host RAM -- and the ollama runner's own resident memory stays inside the allowance below. A rung that spills is printed as `does not fit`, with its per-card memory, and the ladder stops there: a partial residency is a different measurement and is never reported as the card's window.", "pcie_note": "This card negotiated PCIe x16 of a x16-capable generation-3 link -- the same x16 slot the 3090 used, and the slot on which the pair bench measured a x16 + x4 asymmetry. With one card there is no per-token traffic BETWEEN cards, so the link is not in the decode path the way it was for the split pair; it still carries the model in and the tokens out. Every run records `pcie.link.width.current` UNDER LOAD, because an idle card drops its link to save power and a width read at rest would flatter the result.", "thermal_layout_note": "ONE GeForce RTX 3090 Ti 24 GB in a consumer desktop, alone on the board's x16 slot -- the same box, the same case and the same airflow in which one GeForce RTX 3090 24 GB was measured earlier the same day and two GeForce RTX 3080 10 GB cards the night before. A temperature here is a reading of THAT arrangement. Against the single 3090 the arrangement is the same one, which is what makes the two cards' thermal rows comparable; against the PAIR it is not, because with one card there is no second board warming the air or blocking a face. The card is traced on every run.", "card_labels": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": "index 0 - EVGA - GPU-1809d1e1" }, "card_identity": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": 0, "bus_id": "00000000:01:00.0", "pci_sub_device_id": "0x49853842", "vendor": "EVGA", "board_per_operator": "EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB", "vbios": "94.02.A0.00.5C", "serial": null, "pcie_link_width_max": 16, "pcie_link_width_negotiated": 16, "memory_total_mib": 24564, "memory_bus_width_bits": 384, "clocks_max_sm_mhz": 2115, "clocks_max_memory_mhz": 10501, "power_default_limit_w": 450.0, "power_min_limit_w": 100.0, "power_max_limit_w": 480.0, "power_limit_at_bench_start_w": 300.0, "note": "the single card seated in this box's x16 slot -- the same slot the 3090 and the pair of 3080s were measured in. It came up at its own 450 W DEFAULT: the box's boot-time GPU power-cap unit (/etc/desktop/gpu-power-caps.conf, written 2026-08-27 for the two 3090s this box carried then) FAILED BOTH of its rows this boot -- neither UUID it names is in the machine any more -- and its own read-back printed `0, 00000000:01:00.0, 450.00 W, 450.00 W`. The operator then enabled persistence and set 300 W by hand, and the read-back is recorded per arm." } }, "memory_temp_support": { "checked_utc": "2026-09-17T20:02:59Z", "paths": { "nvidia-smi --query-gpu=temperature.memory": { "raw": "0, GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4, N/A", "readable": false }, "nvidia-smi -q -d TEMPERATURE": { "lines": [ "GPU Shutdown Temp : 97 C", "GPU Slowdown Temp : 94 C", "GPU Target Temperature : 83 C", "Memory Current Temp : N/A", "Memory Max Operating Temp : N/A" ], "readable": false }, "NVML NVML_FI_DEV_MEMORY_TEMP": { "available": true, "field_id": 82, "cards": { "0": { "call_rc": 0, "field_rc": 3, "supported": false, "not_supported": true, "value_c": null } } } }, "readable": false, "verdict": "the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead." }, "runner_rss_allowance_gib": 4.0, "num_gpu_option": null, "num_gpu_policy": "layers left to ollama's own planner (`auto`)", "num_gpu_note": "ollama decides for itself how many of a model's layers to put on the card. On the pair of 3080s that decision was measurably conservative -- gemma4:26b loaded 75.2% into VRAM and spilled 4.25 GiB to host RAM while about 5 GiB of the two cards' 20 GiB sat unused -- so the pair bench reported `auto` and a forced layer count side by side. On ONE 24 GB card there is no placement decision to make: the expected reading is 100% on the card with no spill. This harness records `options.num_gpu` on every run as `num_gpu_option` and reads the planner's actual placement back from /api/ps, so a spill on a card with room is visible as a finding rather than absorbed into a tok/s figure. A value at or above the model's layer count means every layer.", "instance_env": { "written_utc": "2026-09-17T20:02:51Z", "unit": "bench-3090ti-one", "shape": "one", "port": 11470, "base_url": "http://127.0.0.1:11470", "cuda_visible_devices": "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4", "OLLAMA_FLASH_ATTENTION": "1", "OLLAMA_KV_CACHE_TYPE": "q8_0", "OLLAMA_CONTEXT_LENGTH": 32768, "OLLAMA_": 1, "OLLAMA_MAX_LOADED_MODELS": 1, "OLLAMA_KEEP_ALIVE": "0 (the server default; every bench request sends its own keep_alive, which wins, and every arm unloads on the way out)", "models_dir": "/usr/share/ollama/.ollama/models", "models_dir_writable_by_this_user": false, "api_version": { "version": "0.32.13" }, "power_limits_at_start_w": "0, 350.00 W;", "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 16;" }, "ups_before_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-17T20:02:59Z", "available": false, "error": "upsc rc=1: Error: Connection failure: Connection refused" }, "levels": [ { "num_ctx": 4096, "started_utc": "2026-09-17T20:02:59Z", "loads": true, "size_total": 17201018960, "size_vram": 17201018960, "context_length_loaded": 4096, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17926.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 304.0 }, "runner_peak_rss_gib": 1.28, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 52100, "rss_kb": 1347316, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3488 457 43 43684 43552\nSwap: 24575 1 24574", "kv_footprint_bytes": 0, "kv_footprint_gib": 0.0, "runs": [ { "ttft_ms": 579.87, "first_token_was_thinking": false, "wall_s": 2.3107, "eval_count": 256, "eval_duration_ns": 1727671000, "decode_tok_s": 148.176, "prompt_eval_count": 536, "prompt_eval_duration_ns": 115792000, "prefill_tok_s": 4628.99, "load_duration_ns": 459462832, "total_duration_ns": 2307327959, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.325, "decode_tok_s_wall_gross": 110.787, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 96.05, "median": 96.31, "max": 96.41, "mean": 96.26, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 96.26 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "power_w_all_cards": 96.26, "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": 5, "mean_w": 267.99, "max_w": 322.4, "median_w": 299.67, "mean_w_whole_window": 194.3, "max_w_whole_window": 322.4, "limit_w": 350.0, "over_idle_w": 171.73, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 96.0, "median": 188.24, "max": 322.4, "mean": 194.3, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 92.0, "mean": 50.9, "n": 9 }, "temperature_c": { "min": 40.0, "median": 45.0, "max": 46.0, "mean": 43.8, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2055.0, "mean": 2001.7, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 188.24, "median": 299.67, "max": 322.4, "mean": 267.99, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 92.0, "max": 92.0, "mean": 91.6, "n": 5 }, "temperature_c": { "min": 45.0, "median": 46.0, "max": 46.0, "mean": 45.8, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1808.59, "power_mean_w_all_cards": 267.99, "energy_j_per_1k_tokens_all_cards": 1808.59, "watts_per_tok_s": 1.8086, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 46.0, "temp_min_c": 45.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 267.99, "power_pct_of_limit": 76.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 46.0, "temp_min_c": 45.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 267.99, "power_pct_of_limit": 76.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 543.47, "first_token_was_thinking": false, "wall_s": 2.2766, "eval_count": 256, "eval_duration_ns": 1730048000, "decode_tok_s": 147.973, "prompt_eval_count": 536, "prompt_eval_duration_ns": 87583000, "prefill_tok_s": 6119.909, "load_duration_ns": 451690839, "total_duration_ns": 2273266911, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.136, "decode_tok_s_wall_gross": 112.45, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 16.0, "median": 16.43, "max": 16.5, "mean": 16.31, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16.31 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "power_w_all_cards": 16.31, "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": 6, "mean_w": 226.58, "max_w": 323.15, "median_w": 252.93, "mean_w_whole_window": 157.27, "max_w_whole_window": 323.15, "limit_w": 350.0, "over_idle_w": 210.27, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 16.12, "median": 156.48, "max": 323.15, "mean": 157.27, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 45.0, "max": 92.0, "mean": 45.7, "n": 9 }, "temperature_c": { "min": 39.0, "median": 47.0, "max": 47.0, "mean": 44.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2040.0, "max": 2040.0, "mean": 1620.0, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 51.51, "median": 252.93, "max": 323.15, "mean": 226.58, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 6 }, "utilization_pct": { "min": 45.0, "median": 68.5, "max": 92.0, "mean": 68.5, "n": 6 }, "temperature_c": { "min": 45.0, "median": 47.0, "max": 47.0, "mean": 46.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1531.23, "power_mean_w_all_cards": 226.58, "energy_j_per_1k_tokens_all_cards": 1531.23, "watts_per_tok_s": 1.5312, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 47.0, "temp_min_c": 45.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 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": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 226.58, "power_pct_of_limit": 64.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 47.0, "temp_min_c": 45.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 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": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 226.58, "power_pct_of_limit": 64.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 537.97, "first_token_was_thinking": false, "wall_s": 2.2698, "eval_count": 256, "eval_duration_ns": 1728695000, "decode_tok_s": 148.089, "prompt_eval_count": 536, "prompt_eval_duration_ns": 78994000, "prefill_tok_s": 6785.325, "load_duration_ns": 454634169, "total_duration_ns": 2266452923, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.243, "decode_tok_s_wall_gross": 112.785, "num_ctx_option": 4096, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 97.18, "median": 97.19, "max": 97.36, "mean": 97.24, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 97.24 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "power_w_all_cards": 97.24, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 4, "mean_w": 293.76, "max_w": 324.05, "median_w": 314.7, "mean_w_whole_window": 199.17, "max_w_whole_window": 324.05, "limit_w": 350.0, "over_idle_w": 196.52, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 96.87, "median": 202.89, "max": 324.05, "mean": 199.17, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 92.0, "mean": 40.9, "n": 9 }, "temperature_c": { "min": 43.0, "median": 48.0, "max": 49.0, "mean": 46.7, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 221.59, "median": 314.7, "max": 324.05, "mean": 293.76, "n": 4 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18230.0, "median": 18230.0, "max": 18230.0, "mean": 18230.0, "n": 4 }, "utilization_pct": { "min": 92.0, "median": 92.0, "max": 92.0, "mean": 92.0, "n": 4 }, "temperature_c": { "min": 48.0, "median": 49.0, "max": 49.0, "mean": 48.8, "n": 4 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 4 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 4 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 4 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 4 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 4 }, "busy_samples": 4 } } }, "energy_j_per_1k_tokens": 1983.68, "power_mean_w_all_cards": 293.76, "energy_j_per_1k_tokens_all_cards": 1983.68, "watts_per_tok_s": 1.9837, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 49.0, "temp_min_c": 48.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 293.76, "power_pct_of_limit": 83.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 49.0, "temp_min_c": 48.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 293.76, "power_pct_of_limit": 83.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.973, "median": 148.089, "max": 148.176, "mean": 148.079, "n": 3 }, "ttft_ms": { "min": 537.97, "median": 543.47, "max": 579.87, "mean": 553.77, "n": 3 }, "power_mean_w": { "min": 226.58, "median": 267.99, "max": 293.76, "mean": 262.78, "n": 3 }, "power_mean_w_all_cards": { "min": 226.58, "median": 267.99, "max": 293.76, "mean": 262.78, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1531.23, "median": 1808.59, "max": 1983.68, "mean": 1774.5, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1531.23, "median": 1808.59, "max": 1983.68, "mean": 1774.5, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 46.0, "median": 47.0, "max": 49.0, "mean": 47.3, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 } }, "finished_utc": "2026-09-17T20:04:03Z" }, { "num_ctx": 8192, "started_utc": "2026-09-17T20:04:03Z", "loads": true, "size_total": 17372607938, "size_vram": 17372607938, "context_length_loaded": 8192, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 208.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18230.0 }, "runner_peak_rss_gib": 1.29, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 52759, "rss_kb": 1357388, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3543 389 69 43722 43497\nSwap: 24575 1 24574", "kv_footprint_bytes": 171588978, "kv_footprint_gib": 0.16, "runs": [ { "ttft_ms": 557.32, "first_token_was_thinking": false, "wall_s": 2.2976, "eval_count": 256, "eval_duration_ns": 1737057000, "decode_tok_s": 147.376, "prompt_eval_count": 536, "prompt_eval_duration_ns": 97964000, "prefill_tok_s": 5471.398, "load_duration_ns": 455083785, "total_duration_ns": 2294154325, "streamed_chunks": 256, "response_chars": 1057, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.529, "decode_tok_s_wall_gross": 111.421, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 16.15, "median": 16.19, "max": 16.35, "mean": 16.23, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16.23 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "power_w_all_cards": 16.23, "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": 5, "mean_w": 261.45, "max_w": 324.36, "median_w": 318.36, "mean_w_whole_window": 158.04, "max_w_whole_window": 324.36, "limit_w": 350.0, "over_idle_w": 245.22, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 16.12, "median": 151.79, "max": 324.36, "mean": 158.04, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 41.0, "median": 49.0, "max": 50.0, "mean": 46.6, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2040.0, "max": 2040.0, "mean": 1620.0, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 151.79, "median": 318.36, "max": 324.36, "mean": 261.45, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 49.0, "median": 49.0, "max": 50.0, "mean": 49.4, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1774.04, "power_mean_w_all_cards": 261.45, "energy_j_per_1k_tokens_all_cards": 1774.04, "watts_per_tok_s": 1.774, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 50.0, "temp_min_c": 49.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 261.45, "power_pct_of_limit": 74.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 50.0, "temp_min_c": 49.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 261.45, "power_pct_of_limit": 74.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 534.93, "first_token_was_thinking": false, "wall_s": 2.2706, "eval_count": 256, "eval_duration_ns": 1732687000, "decode_tok_s": 147.747, "prompt_eval_count": 536, "prompt_eval_duration_ns": 79647000, "prefill_tok_s": 6729.695, "load_duration_ns": 451010177, "total_duration_ns": 2267209847, "streamed_chunks": 256, "response_chars": 1057, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.917, "decode_tok_s_wall_gross": 112.745, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.11, "median": 99.2, "max": 99.39, "mean": 99.23, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.23 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "power_w_all_cards": 99.23, "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": 6, "mean_w": 252.39, "max_w": 325.05, "median_w": 272.31, "mean_w_whole_window": 201.48, "max_w_whole_window": 325.05, "limit_w": 350.0, "over_idle_w": 153.16, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.05, "median": 198.37, "max": 325.05, "mean": 201.48, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 87.0, "max": 91.0, "mean": 59.8, "n": 9 }, "temperature_c": { "min": 45.0, "median": 50.0, "max": 51.0, "mean": 48.8, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2055.0, "mean": 2005.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 122.08, "median": 272.31, "max": 325.05, "mean": 252.39, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 6 }, "utilization_pct": { "min": 87.0, "median": 91.0, "max": 91.0, "mean": 89.7, "n": 6 }, "temperature_c": { "min": 50.0, "median": 51.0, "max": 51.0, "mean": 50.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2047.5, "max": 2055.0, "mean": 2047.5, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1708.25, "power_mean_w_all_cards": 252.39, "energy_j_per_1k_tokens_all_cards": 1708.25, "watts_per_tok_s": 1.7083, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 51.0, "temp_min_c": 50.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2047.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.39, "power_pct_of_limit": 72.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 51.0, "temp_min_c": 50.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2047.5, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.39, "power_pct_of_limit": 72.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 546.86, "first_token_was_thinking": false, "wall_s": 2.2807, "eval_count": 256, "eval_duration_ns": 1730851000, "decode_tok_s": 147.904, "prompt_eval_count": 536, "prompt_eval_duration_ns": 88064000, "prefill_tok_s": 6086.483, "load_duration_ns": 454220664, "total_duration_ns": 2277372663, "streamed_chunks": 256, "response_chars": 1057, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.07, "decode_tok_s_wall_gross": 112.245, "num_ctx_option": 8192, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 17.01, "median": 17.18, "max": 17.29, "mean": 17.16, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17.16 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "power_w_all_cards": 17.16, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 264.47, "max_w": 325.42, "median_w": 320.32, "mean_w_whole_window": 159.18, "max_w_whole_window": 325.42, "limit_w": 350.0, "over_idle_w": 247.31, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 16.84, "median": 157.97, "max": 325.42, "mean": 159.18, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 92.0, "mean": 50.8, "n": 9 }, "temperature_c": { "min": 43.0, "median": 51.0, "max": 52.0, "mean": 48.4, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2055.0, "max": 2055.0, "mean": 1628.3, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 157.97, "median": 320.32, "max": 325.42, "mean": 264.47, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18438.0, "median": 18438.0, "max": 18438.0, "mean": 18438.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 92.0, "mean": 91.4, "n": 5 }, "temperature_c": { "min": 51.0, "median": 51.0, "max": 52.0, "mean": 51.4, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1788.12, "power_mean_w_all_cards": 264.47, "energy_j_per_1k_tokens_all_cards": 1788.12, "watts_per_tok_s": 1.7881, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 52.0, "temp_min_c": 51.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 264.47, "power_pct_of_limit": 75.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 52.0, "temp_min_c": 51.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 264.47, "power_pct_of_limit": 75.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.376, "median": 147.747, "max": 147.904, "mean": 147.676, "n": 3 }, "ttft_ms": { "min": 534.93, "median": 546.86, "max": 557.32, "mean": 546.37, "n": 3 }, "power_mean_w": { "min": 252.39, "median": 261.45, "max": 264.47, "mean": 259.44, "n": 3 }, "power_mean_w_all_cards": { "min": 252.39, "median": 261.45, "max": 264.47, "mean": 259.44, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1708.25, "median": 1774.04, "max": 1788.12, "mean": 1756.8, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1708.25, "median": 1774.04, "max": 1788.12, "mean": 1756.8, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 50.0, "median": 51.0, "max": 52.0, "mean": 51.0, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 } }, "finished_utc": "2026-09-17T20:05:24Z" }, { "num_ctx": 16384, "started_utc": "2026-09-17T20:05:24Z", "loads": true, "size_total": 17239291984, "size_vram": 17239291984, "context_length_loaded": 16384, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": -46.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18438.0 }, "runner_peak_rss_gib": 1.3, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 53803, "rss_kb": 1360412, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3534 558 56 43549 43506\nSwap: 24575 1 24574", "kv_footprint_bytes": 38273024, "kv_footprint_gib": 0.04, "runs": [ { "ttft_ms": 557.56, "first_token_was_thinking": false, "wall_s": 2.2997, "eval_count": 256, "eval_duration_ns": 1739301000, "decode_tok_s": 147.186, "prompt_eval_count": 536, "prompt_eval_duration_ns": 101540000, "prefill_tok_s": 5278.708, "load_duration_ns": 451694372, "total_duration_ns": 2296465683, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.371, "decode_tok_s_wall_gross": 111.319, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 17.75, "median": 17.94, "max": 18.21, "mean": 17.97, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 17.97 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "power_w_all_cards": 17.97, "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": 6, "mean_w": 224.83, "max_w": 328.0, "median_w": 241.92, "mean_w_whole_window": 155.98, "max_w_whole_window": 328.0, "limit_w": 350.0, "over_idle_w": 206.86, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 17.64, "median": 153.95, "max": 328.0, "mean": 155.98, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 26.0, "max": 91.0, "mean": 39.0, "n": 9 }, "temperature_c": { "min": 44.0, "median": 52.0, "max": 53.0, "mean": 49.7, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2055.0, "max": 2055.0, "mean": 1628.3, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 55.3, "median": 241.92, "max": 328.0, "mean": 224.83, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 6 }, "utilization_pct": { "min": 26.0, "median": 58.5, "max": 91.0, "mean": 58.5, "n": 6 }, "temperature_c": { "min": 51.0, "median": 52.0, "max": 53.0, "mean": 52.2, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2055.0, "max": 2055.0, "mean": 2052.5, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1527.53, "power_mean_w_all_cards": 224.83, "energy_j_per_1k_tokens_all_cards": 1527.53, "watts_per_tok_s": 1.5275, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 53.0, "temp_min_c": 51.0, "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": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 224.83, "power_pct_of_limit": 64.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 53.0, "temp_min_c": 51.0, "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": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 224.83, "power_pct_of_limit": 64.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 538.36, "first_token_was_thinking": false, "wall_s": 2.2763, "eval_count": 256, "eval_duration_ns": 1734830000, "decode_tok_s": 147.565, "prompt_eval_count": 536, "prompt_eval_duration_ns": 80887000, "prefill_tok_s": 6626.528, "load_duration_ns": 453186813, "total_duration_ns": 2272868761, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.727, "decode_tok_s_wall_gross": 112.464, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 99.76, "median": 99.93, "max": 100.03, "mean": 99.91, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 99.91 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "power_w_all_cards": 99.91, "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": 4, "mean_w": 299.47, "max_w": 329.5, "median_w": 320.75, "mean_w_whole_window": 203.7, "max_w_whole_window": 329.5, "limit_w": 350.0, "over_idle_w": 199.56, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 99.63, "median": 208.27, "max": 329.5, "mean": 203.7, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 91.0, "mean": 40.4, "n": 9 }, "temperature_c": { "min": 48.0, "median": 53.0, "max": 54.0, "mean": 51.6, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2055.0, "max": 2055.0, "mean": 2010.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 226.86, "median": 320.75, "max": 329.5, "mean": 299.47, "n": 4 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 4 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 4 }, "temperature_c": { "min": 53.0, "median": 53.5, "max": 54.0, "mean": 53.5, "n": 4 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 4 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 4 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 4 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 4 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 4 }, "busy_samples": 4 } } }, "energy_j_per_1k_tokens": 2029.41, "power_mean_w_all_cards": 299.47, "energy_j_per_1k_tokens_all_cards": 2029.41, "watts_per_tok_s": 2.0294, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 54.0, "temp_min_c": 53.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 299.47, "power_pct_of_limit": 85.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 54.0, "temp_min_c": 53.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 299.47, "power_pct_of_limit": 85.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 552.26, "first_token_was_thinking": false, "wall_s": 2.2907, "eval_count": 256, "eval_duration_ns": 1735277000, "decode_tok_s": 147.527, "prompt_eval_count": 536, "prompt_eval_duration_ns": 89521000, "prefill_tok_s": 5987.422, "load_duration_ns": 458471560, "total_duration_ns": 2287314274, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.686, "decode_tok_s_wall_gross": 111.758, "num_ctx_option": 16384, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 16.69, "median": 16.75, "max": 16.81, "mean": 16.75, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 16.75 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "power_w_all_cards": 16.75, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 7, "mean_w": 194.8, "max_w": 328.62, "median_w": 177.71, "mean_w_whole_window": 155.2, "max_w_whole_window": 328.62, "limit_w": 350.0, "over_idle_w": 178.05, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 16.59, "median": 152.6, "max": 328.62, "mean": 155.2, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 55.0, "n": 9 }, "temperature_c": { "min": 45.0, "median": 53.0, "max": 54.0, "mean": 50.4, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2055.0, "max": 2055.0, "mean": 1615.0, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 18.51, "median": 177.71, "max": 328.62, "mean": 194.8, "n": 7 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18392.0, "median": 18392.0, "max": 18392.0, "mean": 18392.0, "n": 7 }, "utilization_pct": { "min": 20.0, "median": 91.0, "max": 91.0, "mean": 70.7, "n": 7 }, "temperature_c": { "min": 46.0, "median": 53.0, "max": 54.0, "mean": 52.0, "n": 7 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 7 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 7 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2055.0, "max": 2055.0, "mean": 2016.4, "n": 7 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 7 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 7 }, "busy_samples": 7 } } }, "energy_j_per_1k_tokens": 1320.44, "power_mean_w_all_cards": 194.8, "energy_j_per_1k_tokens_all_cards": 1320.44, "watts_per_tok_s": 1.3204, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 7, "temp_max_c": 54.0, "temp_min_c": 46.0, "temp_rise_c": 8.0, "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": 1920.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 194.8, "power_pct_of_limit": 55.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "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 55.7% of the cap and the card at 54 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-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 7, "temp_max_c": 54.0, "temp_min_c": 46.0, "temp_rise_c": 8.0, "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": 1920.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": true, "power_limit_w": 350.0, "power_mean_w": 194.8, "power_pct_of_limit": 55.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "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 55.7% of the cap and the card at 54 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)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.186, "median": 147.527, "max": 147.565, "mean": 147.426, "n": 3 }, "ttft_ms": { "min": 538.36, "median": 552.26, "max": 557.56, "mean": 549.39, "n": 3 }, "power_mean_w": { "min": 194.8, "median": 224.83, "max": 299.47, "mean": 239.7, "n": 3 }, "power_mean_w_all_cards": { "min": 194.8, "median": 224.83, "max": 299.47, "mean": 239.7, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1320.44, "median": 1527.53, "max": 2029.41, "mean": 1625.79, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1320.44, "median": 1527.53, "max": 2029.41, "mean": 1625.79, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 53.0, "median": 54.0, "max": 54.0, "mean": 53.7, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 } }, "finished_utc": "2026-09-17T20:06:45Z" }, { "num_ctx": 32768, "started_utc": "2026-09-17T20:06:45Z", "loads": true, "size_total": 17323178064, "size_vram": 17323178064, "context_length_loaded": 32768, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 250.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18392.0 }, "runner_peak_rss_gib": 1.32, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 54840, "rss_kb": 1378908, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3635 456 72 43565 43405\nSwap: 24575 1 24574", "kv_footprint_bytes": 122159104, "kv_footprint_gib": 0.11, "runs": [ { "ttft_ms": 548.81, "first_token_was_thinking": false, "wall_s": 2.2744, "eval_count": 256, "eval_duration_ns": 1722811000, "decode_tok_s": 148.594, "prompt_eval_count": 536, "prompt_eval_duration_ns": 91077000, "prefill_tok_s": 5885.13, "load_duration_ns": 453453247, "total_duration_ns": 2271236730, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.774, "decode_tok_s_wall_gross": 112.556, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 100.0, "median": 100.19, "max": 100.25, "mean": 100.15, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 100.15 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "power_w_all_cards": 100.15, "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": 6, "mean_w": 255.25, "max_w": 330.5, "median_w": 275.19, "mean_w_whole_window": 203.65, "max_w_whole_window": 330.5, "limit_w": 350.0, "over_idle_w": 155.1, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 100.22, "median": 198.66, "max": 330.5, "mean": 203.65, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 86.0, "max": 91.0, "mean": 59.6, "n": 9 }, "temperature_c": { "min": 49.0, "median": 54.0, "max": 55.0, "mean": 52.8, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2055.0, "max": 2055.0, "mean": 2010.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 122.26, "median": 275.19, "max": 330.5, "mean": 255.25, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 6 }, "utilization_pct": { "min": 86.0, "median": 91.0, "max": 91.0, "mean": 89.3, "n": 6 }, "temperature_c": { "min": 54.0, "median": 55.0, "max": 55.0, "mean": 54.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1717.76, "power_mean_w_all_cards": 255.25, "energy_j_per_1k_tokens_all_cards": 1717.76, "watts_per_tok_s": 1.7178, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 55.0, "temp_min_c": 54.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 255.25, "power_pct_of_limit": 72.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 55.0, "temp_min_c": 54.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 255.25, "power_pct_of_limit": 72.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 540.56, "first_token_was_thinking": false, "wall_s": 2.2727, "eval_count": 256, "eval_duration_ns": 1729245000, "decode_tok_s": 148.041, "prompt_eval_count": 536, "prompt_eval_duration_ns": 79027000, "prefill_tok_s": 6782.492, "load_duration_ns": 457202493, "total_duration_ns": 2269457462, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.217, "decode_tok_s_wall_gross": 112.641, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 100.92, "median": 101.12, "max": 101.28, "mean": 101.11, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 101.11 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "power_w_all_cards": 101.11, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 278.75, "max_w": 330.85, "median_w": 306.19, "mean_w_whole_window": 202.7, "max_w_whole_window": 330.85, "limit_w": 350.0, "over_idle_w": 177.64, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 101.51, "median": 198.97, "max": 330.85, "mean": 202.7, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 50.0, "median": 55.0, "max": 56.0, "mean": 53.4, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2055.0, "max": 2055.0, "mean": 2010.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 198.97, "median": 306.19, "max": 330.85, "mean": 278.75, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 55.0, "median": 55.0, "max": 56.0, "mean": 55.4, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1882.92, "power_mean_w_all_cards": 278.75, "energy_j_per_1k_tokens_all_cards": 1882.92, "watts_per_tok_s": 1.8829, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 278.75, "power_pct_of_limit": 79.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 278.75, "power_pct_of_limit": 79.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 552.41, "first_token_was_thinking": false, "wall_s": 2.2835, "eval_count": 256, "eval_duration_ns": 1727874000, "decode_tok_s": 148.159, "prompt_eval_count": 536, "prompt_eval_duration_ns": 89929000, "prefill_tok_s": 5960.258, "load_duration_ns": 458209315, "total_duration_ns": 2280023360, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.306, "decode_tok_s_wall_gross": 112.109, "num_ctx_option": 32768, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 17.89, "median": 18.23, "max": 18.41, "mean": 18.18, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18.18 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "power_w_all_cards": 18.18, "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": 6, "mean_w": 232.91, "max_w": 331.17, "median_w": 256.31, "mean_w_whole_window": 162.29, "max_w_whole_window": 331.17, "limit_w": 350.0, "over_idle_w": 214.73, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 18.29, "median": 166.33, "max": 331.17, "mean": 162.29, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 45.0, "max": 91.0, "mean": 45.3, "n": 9 }, "temperature_c": { "min": 48.0, "median": 55.0, "max": 56.0, "mean": 53.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2055.0, "max": 2055.0, "mean": 1630.0, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 56.82, "median": 256.31, "max": 331.17, "mean": 232.91, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18642.0, "median": 18642.0, "max": 18642.0, "mean": 18642.0, "n": 6 }, "utilization_pct": { "min": 45.0, "median": 68.0, "max": 91.0, "mean": 68.0, "n": 6 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 56.0, "mean": 55.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1572.03, "power_mean_w_all_cards": 232.91, "energy_j_per_1k_tokens_all_cards": 1572.03, "watts_per_tok_s": 1.572, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 232.91, "power_pct_of_limit": 66.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 232.91, "power_pct_of_limit": 66.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 148.041, "median": 148.159, "max": 148.594, "mean": 148.265, "n": 3 }, "ttft_ms": { "min": 540.56, "median": 548.81, "max": 552.41, "mean": 547.26, "n": 3 }, "power_mean_w": { "min": 232.91, "median": 255.25, "max": 278.75, "mean": 255.64, "n": 3 }, "power_mean_w_all_cards": { "min": 232.91, "median": 255.25, "max": 278.75, "mean": 255.64, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1572.03, "median": 1717.76, "max": 1882.92, "mean": 1724.24, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1572.03, "median": 1717.76, "max": 1882.92, "mean": 1724.24, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 55.0, "median": 56.0, "max": 56.0, "mean": 55.7, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 } }, "finished_utc": "2026-09-17T20:07:46Z" }, { "num_ctx": 49152, "started_utc": "2026-09-17T20:07:46Z", "loads": true, "size_total": 17407064144, "size_vram": 17407064144, "context_length_loaded": 49152, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 250.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18642.0 }, "runner_peak_rss_gib": 1.33, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 55680, "rss_kb": 1395984, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3615 635 88 43422 43425\nSwap: 24575 1 24574", "kv_footprint_bytes": 206045184, "kv_footprint_gib": 0.19, "runs": [ { "ttft_ms": 552.24, "first_token_was_thinking": false, "wall_s": 2.2925, "eval_count": 256, "eval_duration_ns": 1737403000, "decode_tok_s": 147.346, "prompt_eval_count": 536, "prompt_eval_duration_ns": 91367000, "prefill_tok_s": 5866.451, "load_duration_ns": 456436882, "total_duration_ns": 2289276090, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.53, "decode_tok_s_wall_gross": 111.668, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.0, "median": 102.22, "max": 102.47, "mean": 102.23, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.23 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "power_w_all_cards": 102.23, "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": 6, "mean_w": 252.29, "max_w": 329.6, "median_w": 267.19, "mean_w_whole_window": 202.37, "max_w_whole_window": 329.6, "limit_w": 350.0, "over_idle_w": 150.06, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.49, "median": 190.67, "max": 329.6, "mean": 202.37, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 92.0, "mean": 60.9, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 57.0, "mean": 55.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2055.0, "mean": 2001.7, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 129.8, "median": 267.19, "max": 329.6, "mean": 252.29, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 6 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 92.0, "mean": 91.3, "n": 6 }, "temperature_c": { "min": 55.0, "median": 57.0, "max": 57.0, "mean": 56.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2055.0, "mean": 2042.5, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1712.22, "power_mean_w_all_cards": 252.29, "energy_j_per_1k_tokens_all_cards": 1712.22, "watts_per_tok_s": 1.7122, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.29, "power_pct_of_limit": 72.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.29, "power_pct_of_limit": 72.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 543.18, "first_token_was_thinking": false, "wall_s": 2.2742, "eval_count": 256, "eval_duration_ns": 1728043000, "decode_tok_s": 148.144, "prompt_eval_count": 536, "prompt_eval_duration_ns": 79567000, "prefill_tok_s": 6736.461, "load_duration_ns": 459039296, "total_duration_ns": 2270851586, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.313, "decode_tok_s_wall_gross": 112.568, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.63, "median": 102.8, "max": 102.81, "mean": 102.75, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.75 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "power_w_all_cards": 102.75, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 283.55, "max_w": 333.83, "median_w": 323.89, "mean_w_whole_window": 206.69, "max_w_whole_window": 333.83, "limit_w": 350.0, "over_idle_w": 180.8, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.57, "median": 198.81, "max": 333.83, "mean": 206.69, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 92.0, "mean": 50.9, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 58.0, "mean": 55.8, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 198.81, "median": 323.89, "max": 333.83, "mean": 283.55, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 92.0, "max": 92.0, "mean": 91.6, "n": 5 }, "temperature_c": { "min": 57.0, "median": 58.0, "max": 58.0, "mean": 57.6, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1914.01, "power_mean_w_all_cards": 283.55, "energy_j_per_1k_tokens_all_cards": 1914.01, "watts_per_tok_s": 1.914, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 58.0, "temp_min_c": 57.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 283.55, "power_pct_of_limit": 81.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 58.0, "temp_min_c": 57.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 283.55, "power_pct_of_limit": 81.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 540.85, "first_token_was_thinking": false, "wall_s": 2.2815, "eval_count": 256, "eval_duration_ns": 1737457000, "decode_tok_s": 147.342, "prompt_eval_count": 536, "prompt_eval_duration_ns": 80559000, "prefill_tok_s": 6653.509, "load_duration_ns": 455683323, "total_duration_ns": 2277992685, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.497, "decode_tok_s_wall_gross": 112.207, "num_ctx_option": 49152, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 103.02, "median": 103.03, "max": 103.27, "mean": 103.11, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.11 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "power_w_all_cards": 103.11, "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": 6, "mean_w": 255.3, "max_w": 332.0, "median_w": 274.0, "mean_w_whole_window": 204.66, "max_w_whole_window": 332.0, "limit_w": 350.0, "over_idle_w": 152.19, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.09, "median": 197.28, "max": 332.0, "mean": 204.66, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 86.0, "max": 91.0, "mean": 59.6, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 57.0, "mean": 55.3, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 50.0, "median": 50.0, "max": 50.0, "mean": 50.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 122.7, "median": 274.0, "max": 332.0, "mean": 255.3, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 18892.0, "median": 18892.0, "max": 18892.0, "mean": 18892.0, "n": 6 }, "utilization_pct": { "min": 86.0, "median": 91.0, "max": 91.0, "mean": 89.3, "n": 6 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.8, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 50.0, "median": 50.0, "max": 50.0, "mean": 50.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1732.71, "power_mean_w_all_cards": 255.3, "energy_j_per_1k_tokens_all_cards": 1732.71, "watts_per_tok_s": 1.7327, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 50.0, "fan_max_pct": 50.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 255.3, "power_pct_of_limit": 72.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 50.0, "fan_max_pct": 50.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 255.3, "power_pct_of_limit": 72.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.342, "median": 147.346, "max": 148.144, "mean": 147.611, "n": 3 }, "ttft_ms": { "min": 540.85, "median": 543.18, "max": 552.24, "mean": 545.42, "n": 3 }, "power_mean_w": { "min": 252.29, "median": 255.3, "max": 283.55, "mean": 263.71, "n": 3 }, "power_mean_w_all_cards": { "min": 252.29, "median": 255.3, "max": 283.55, "mean": 263.71, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1712.22, "median": 1732.71, "max": 1914.01, "mean": 1786.31, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1712.22, "median": 1732.71, "max": 1914.01, "mean": 1786.31, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 58.0, "mean": 57.3, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 50.0, "mean": 16.7, "n": 3 } }, "finished_utc": "2026-09-17T20:08:37Z" }, { "num_ctx": 65536, "started_utc": "2026-09-17T20:08:37Z", "loads": true, "size_total": 17490950224, "size_vram": 17490950224, "context_length_loaded": 65536, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 250.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18892.0 }, "runner_peak_rss_gib": 1.35, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 56461, "rss_kb": 1413108, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3667 575 104 43446 43373\nSwap: 24575 1 24574", "kv_footprint_bytes": 289931264, "kv_footprint_gib": 0.27, "runs": [ { "ttft_ms": 549.03, "first_token_was_thinking": false, "wall_s": 2.2871, "eval_count": 256, "eval_duration_ns": 1735000000, "decode_tok_s": 147.55, "prompt_eval_count": 536, "prompt_eval_duration_ns": 90771000, "prefill_tok_s": 5904.97, "load_duration_ns": 453990068, "total_duration_ns": 2283684687, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.712, "decode_tok_s_wall_gross": 111.931, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.71, "median": 102.78, "max": 102.84, "mean": 102.78, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.78 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "power_w_all_cards": 102.78, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 274.65, "max_w": 330.88, "median_w": 305.59, "mean_w_whole_window": 200.7, "max_w_whole_window": 330.88, "limit_w": 350.0, "over_idle_w": 171.87, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.87, "median": 193.36, "max": 330.88, "mean": 200.7, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 51.0, "median": 56.0, "max": 57.0, "mean": 54.6, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 49.0, "max": 49.0, "mean": 48.6, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 193.36, "median": 305.59, "max": 330.88, "mean": 274.65, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.6, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 48.0, "max": 49.0, "mean": 48.2, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1861.4, "power_mean_w_all_cards": 274.65, "energy_j_per_1k_tokens_all_cards": 1861.4, "watts_per_tok_s": 1.8614, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.2, "fan_max_pct": 49.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 274.65, "power_pct_of_limit": 78.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.2, "fan_max_pct": 49.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 274.65, "power_pct_of_limit": 78.5, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 547.12, "first_token_was_thinking": false, "wall_s": 2.2828, "eval_count": 256, "eval_duration_ns": 1732632000, "decode_tok_s": 147.752, "prompt_eval_count": 536, "prompt_eval_duration_ns": 88716000, "prefill_tok_s": 6041.751, "load_duration_ns": 454069394, "total_duration_ns": 2279580437, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.914, "decode_tok_s_wall_gross": 112.142, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 18.38, "median": 18.81, "max": 18.88, "mean": 18.69, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18.69 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "power_w_all_cards": 18.69, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 265.78, "max_w": 327.63, "median_w": 306.69, "mean_w_whole_window": 161.16, "max_w_whole_window": 327.63, "limit_w": 350.0, "over_idle_w": 247.09, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 18.2, "median": 166.36, "max": 327.63, "mean": 161.16, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 90.0, "max": 91.0, "mean": 50.3, "n": 9 }, "temperature_c": { "min": 48.0, "median": 55.0, "max": 56.0, "mean": 53.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 38.0, "median": 39.0, "max": 40.0, "mean": 38.8, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2055.0, "max": 2055.0, "mean": 1630.0, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 166.36, "median": 306.69, "max": 327.63, "mean": 265.78, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 5 }, "utilization_pct": { "min": 90.0, "median": 91.0, "max": 91.0, "mean": 90.6, "n": 5 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 56.0, "mean": 55.8, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 38.0, "median": 38.0, "max": 39.0, "mean": 38.2, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2055.0, "median": 2055.0, "max": 2055.0, "mean": 2055.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1798.82, "power_mean_w_all_cards": 265.78, "energy_j_per_1k_tokens_all_cards": 1798.82, "watts_per_tok_s": 1.7988, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 38.2, "fan_max_pct": 39.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 265.78, "power_pct_of_limit": 75.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 56.0, "temp_min_c": 55.0, "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": 38.2, "fan_max_pct": 39.0, "clock_floor_mhz": 2055.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2055.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 265.78, "power_pct_of_limit": 75.9, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 541.87, "first_token_was_thinking": false, "wall_s": 2.2841, "eval_count": 256, "eval_duration_ns": 1739077000, "decode_tok_s": 147.205, "prompt_eval_count": 536, "prompt_eval_duration_ns": 80451000, "prefill_tok_s": 6662.44, "load_duration_ns": 456967841, "total_duration_ns": 2280660804, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.364, "decode_tok_s_wall_gross": 112.079, "num_ctx_option": 65536, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.6, "median": 102.6, "max": 102.67, "mean": 102.62, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 102.62 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "power_w_all_cards": 102.62, "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": 6, "mean_w": 252.64, "max_w": 328.86, "median_w": 264.7, "mean_w_whole_window": 202.64, "max_w_whole_window": 328.86, "limit_w": 350.0, "over_idle_w": 150.02, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 102.54, "median": 203.54, "max": 328.86, "mean": 202.64, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 60.7, "n": 9 }, "temperature_c": { "min": 51.0, "median": 56.0, "max": 57.0, "mean": 54.4, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2055.0, "mean": 2003.3, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 126.49, "median": 264.7, "max": 328.86, "mean": 252.64, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19142.0, "median": 19142.0, "max": 19142.0, "mean": 19142.0, "n": 6 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 6 }, "temperature_c": { "min": 55.0, "median": 56.0, "max": 57.0, "mean": 56.2, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2055.0, "mean": 2045.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1716.25, "power_mean_w_all_cards": 252.64, "energy_j_per_1k_tokens_all_cards": 1716.25, "watts_per_tok_s": 1.7162, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.64, "power_pct_of_limit": 72.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 252.64, "power_pct_of_limit": 72.2, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.205, "median": 147.55, "max": 147.752, "mean": 147.502, "n": 3 }, "ttft_ms": { "min": 541.87, "median": 547.12, "max": 549.03, "mean": 546.01, "n": 3 }, "power_mean_w": { "min": 252.64, "median": 265.78, "max": 274.65, "mean": 264.36, "n": 3 }, "power_mean_w_all_cards": { "min": 252.64, "median": 265.78, "max": 274.65, "mean": 264.36, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1716.25, "median": 1798.82, "max": 1861.4, "mean": 1792.16, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1716.25, "median": 1798.82, "max": 1861.4, "mean": 1792.16, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.7, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 39.0, "max": 49.0, "mean": 29.3, "n": 3 } }, "finished_utc": "2026-09-17T20:09:38Z" }, { "num_ctx": 98304, "started_utc": "2026-09-17T20:09:38Z", "loads": true, "size_total": 17658722384, "size_vram": 17658722384, "context_length_loaded": 98304, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 500.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19142.0 }, "runner_peak_rss_gib": 1.38, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 57145, "rss_kb": 1446280, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3656 524 136 43541 43384\nSwap: 24575 1 24574", "kv_footprint_bytes": 457703424, "kv_footprint_gib": 0.43, "runs": [ { "ttft_ms": 551.99, "first_token_was_thinking": false, "wall_s": 2.2969, "eval_count": 256, "eval_duration_ns": 1741971000, "decode_tok_s": 146.96, "prompt_eval_count": 536, "prompt_eval_duration_ns": 91475000, "prefill_tok_s": 5859.524, "load_duration_ns": 456252332, "total_duration_ns": 2293575358, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.142, "decode_tok_s_wall_gross": 111.456, "num_ctx_option": 98304, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.93, "median": 103.03, "max": 103.16, "mean": 103.04, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.04 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "power_w_all_cards": 103.04, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 275.5, "max_w": 332.25, "median_w": 310.72, "mean_w_whole_window": 201.71, "max_w_whole_window": 332.25, "limit_w": 350.0, "over_idle_w": 172.46, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.32, "median": 186.24, "max": 332.25, "mean": 201.71, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 57.0, "mean": 55.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 1986.7, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 186.24, "median": 310.72, "max": 332.25, "mean": 275.5, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1874.66, "power_mean_w_all_cards": 275.5, "energy_j_per_1k_tokens_all_cards": 1874.66, "watts_per_tok_s": 1.8747, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 57.0, "temp_rise_c": 0.0, "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": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 275.5, "power_pct_of_limit": 78.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 57.0, "temp_rise_c": 0.0, "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": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 275.5, "power_pct_of_limit": 78.7, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 547.69, "first_token_was_thinking": false, "wall_s": 2.2817, "eval_count": 256, "eval_duration_ns": 1731375000, "decode_tok_s": 147.859, "prompt_eval_count": 536, "prompt_eval_duration_ns": 87610000, "prefill_tok_s": 6118.023, "load_duration_ns": 455804211, "total_duration_ns": 2278641467, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.055, "decode_tok_s_wall_gross": 112.196, "num_ctx_option": 98304, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 18.21, "median": 18.28, "max": 18.34, "mean": 18.28, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 18.28 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "power_w_all_cards": 18.28, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 6, "mean_w": 227.51, "max_w": 326.31, "median_w": 247.12, "mean_w_whole_window": 157.93, "max_w_whole_window": 326.31, "limit_w": 350.0, "over_idle_w": 209.23, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 18.07, "median": 160.08, "max": 326.31, "mean": 157.93, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 69.0, "max": 92.0, "mean": 53.7, "n": 9 }, "temperature_c": { "min": 49.0, "median": 57.0, "max": 57.0, "mean": 54.4, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 37.0, "median": 37.0, "max": 38.0, "mean": 37.3, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 210.0, "median": 2040.0, "max": 2055.0, "mean": 1621.7, "n": 9 }, "clocks_mem_mhz": { "min": 405.0, "median": 10251.0, "max": 10251.0, "mean": 8063.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 59.22, "median": 247.12, "max": 326.31, "mean": 227.51, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 6 }, "utilization_pct": { "min": 69.0, "median": 80.5, "max": 92.0, "mean": 80.5, "n": 6 }, "temperature_c": { "min": 56.0, "median": 57.0, "max": 57.0, "mean": 56.8, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 37.0, "median": 37.0, "max": 37.0, "mean": 37.0, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2055.0, "mean": 2042.5, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1538.69, "power_mean_w_all_cards": 227.51, "energy_j_per_1k_tokens_all_cards": 1538.69, "watts_per_tok_s": 1.5387, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 37.0, "fan_max_pct": 37.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 227.51, "power_pct_of_limit": 65.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 37.0, "fan_max_pct": 37.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 227.51, "power_pct_of_limit": 65.0, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 538.53, "first_token_was_thinking": false, "wall_s": 2.264, "eval_count": 256, "eval_duration_ns": 1722548000, "decode_tok_s": 148.617, "prompt_eval_count": 536, "prompt_eval_duration_ns": 79434000, "prefill_tok_s": 6747.74, "load_duration_ns": 454791019, "total_duration_ns": 2260788056, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.788, "decode_tok_s_wall_gross": 113.075, "num_ctx_option": 98304, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 103.0, "median": 103.01, "max": 103.05, "mean": 103.02, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.02 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "power_w_all_cards": 103.02, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 4, "mean_w": 304.88, "max_w": 333.57, "median_w": 328.8, "mean_w_whole_window": 207.12, "max_w_whole_window": 333.57, "limit_w": 350.0, "over_idle_w": 201.86, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.03, "median": 199.84, "max": 333.57, "mean": 207.12, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 92.0, "mean": 40.9, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 58.0, "mean": 55.7, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 228.34, "median": 328.8, "max": 333.57, "mean": 304.88, "n": 4 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 19642.0, "median": 19642.0, "max": 19642.0, "mean": 19642.0, "n": 4 }, "utilization_pct": { "min": 92.0, "median": 92.0, "max": 92.0, "mean": 92.0, "n": 4 }, "temperature_c": { "min": 57.0, "median": 57.5, "max": 58.0, "mean": 57.5, "n": 4 }, "memory_temperature_c": null, "fan_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 4 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 4 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 4 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 4 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 4 }, "busy_samples": 4 } } }, "energy_j_per_1k_tokens": 2051.45, "power_mean_w_all_cards": 304.88, "energy_j_per_1k_tokens_all_cards": 2051.45, "watts_per_tok_s": 2.0514, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 58.0, "temp_min_c": 57.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 304.88, "power_pct_of_limit": 87.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 4, "temp_max_c": 58.0, "temp_min_c": 57.0, "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": 0.0, "fan_max_pct": 0.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 304.88, "power_pct_of_limit": 87.1, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode (the fans never spun up -- zero-RPM idle mode)" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 146.96, "median": 147.859, "max": 148.617, "mean": 147.812, "n": 3 }, "ttft_ms": { "min": 538.53, "median": 547.69, "max": 551.99, "mean": 546.07, "n": 3 }, "power_mean_w": { "min": 227.51, "median": 275.5, "max": 304.88, "mean": 269.3, "n": 3 }, "power_mean_w_all_cards": { "min": 227.51, "median": 275.5, "max": 304.88, "mean": 269.3, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1538.69, "median": 1874.66, "max": 2051.45, "mean": 1821.6, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1538.69, "median": 1874.66, "max": 2051.45, "mean": 1821.6, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 58.0, "mean": 57.3, "n": 3 }, "fan_max_pct": { "min": 0.0, "median": 0.0, "max": 37.0, "mean": 12.3, "n": 3 } }, "finished_utc": "2026-09-17T20:10:39Z" }, { "num_ctx": 131072, "started_utc": "2026-09-17T20:10:39Z", "loads": true, "size_total": 17826494544, "size_vram": 17826494544, "context_length_loaded": 131072, "fits_fully_in_vram": true, "verdict": "fits: fully resident on the card", "offload_split": "100.0% VRAM / 0.0% RAM", "vram_delta_mib": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 500.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "memory_used_mib_per_card_before": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 19642.0 }, "runner_peak_rss_gib": 1.41, "runner_rss_within_allowance": true, "runner_procs": [ { "pid": 57809, "rss_kb": 1480588, "threads": 27 } ], "free_m_at_level": "total used free shared buff/cache available\nMem: 47040 3661 518 168 43574 43379\nSwap: 24575 1 24574", "kv_footprint_bytes": 625475584, "kv_footprint_gib": 0.58, "runs": [ { "ttft_ms": 550.9, "first_token_was_thinking": false, "wall_s": 2.2857, "eval_count": 256, "eval_duration_ns": 1731803000, "decode_tok_s": 147.823, "prompt_eval_count": 536, "prompt_eval_duration_ns": 92399000, "prefill_tok_s": 5800.929, "load_duration_ns": 453832474, "total_duration_ns": 2282276767, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.993, "decode_tok_s_wall_gross": 112.002, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 103.47, "median": 103.48, "max": 103.59, "mean": 103.51, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.51 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "power_w_all_cards": 103.51, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 284.4, "max_w": 332.62, "median_w": 325.85, "mean_w_whole_window": 206.62, "max_w_whole_window": 332.62, "limit_w": 350.0, "over_idle_w": 180.89, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.65, "median": 201.81, "max": 332.62, "mean": 206.62, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 57.0, "mean": 55.2, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 49.0, "median": 49.0, "max": 50.0, "mean": 49.3, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 201.81, "median": 325.85, "max": 332.62, "mean": 284.4, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 49.0, "median": 49.0, "max": 49.0, "mean": 49.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1923.92, "power_mean_w_all_cards": 284.4, "energy_j_per_1k_tokens_all_cards": 1923.92, "watts_per_tok_s": 1.9239, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 57.0, "temp_rise_c": 0.0, "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": 49.0, "fan_max_pct": 49.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 284.4, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 57.0, "temp_rise_c": 0.0, "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": 49.0, "fan_max_pct": 49.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 284.4, "power_pct_of_limit": 81.3, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 538.38, "first_token_was_thinking": false, "wall_s": 2.2746, "eval_count": 256, "eval_duration_ns": 1733351000, "decode_tok_s": 147.691, "prompt_eval_count": 536, "prompt_eval_duration_ns": 79731000, "prefill_tok_s": 6722.605, "load_duration_ns": 454211501, "total_duration_ns": 2271264257, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 146.874, "decode_tok_s_wall_gross": 112.549, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 103.29, "median": 103.34, "max": 103.35, "mean": 103.33, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.33 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "power_w_all_cards": 103.33, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 6, "mean_w": 254.06, "max_w": 331.93, "median_w": 268.1, "mean_w_whole_window": 204.08, "max_w_whole_window": 331.93, "limit_w": 350.0, "over_idle_w": 150.73, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.17, "median": 200.14, "max": 331.93, "mean": 204.08, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 88.0, "max": 92.0, "mean": 60.2, "n": 9 }, "temperature_c": { "min": 52.0, "median": 57.0, "max": 57.0, "mean": 55.1, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 52.0, "max": 52.0, "mean": 51.7, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2040.0, "mean": 2000.0, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 124.54, "median": 268.1, "max": 331.93, "mean": 254.06, "n": 6 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 6 }, "utilization_pct": { "min": 88.0, "median": 91.0, "max": 92.0, "mean": 90.3, "n": 6 }, "temperature_c": { "min": 55.0, "median": 57.0, "max": 57.0, "mean": 56.7, "n": 6 }, "memory_temperature_c": null, "fan_pct": { "min": 51.0, "median": 51.5, "max": 52.0, "mean": 51.5, "n": 6 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 6 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2040.0, "mean": 2040.0, "n": 6 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 6 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 6 }, "busy_samples": 6 } } }, "energy_j_per_1k_tokens": 1720.22, "power_mean_w_all_cards": 254.06, "energy_j_per_1k_tokens_all_cards": 1720.22, "watts_per_tok_s": 1.7202, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 51.5, "fan_max_pct": 52.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 254.06, "power_pct_of_limit": 72.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 6, "temp_max_c": 57.0, "temp_min_c": 55.0, "temp_rise_c": 2.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 51.5, "fan_max_pct": 52.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2040.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 254.06, "power_pct_of_limit": 72.6, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 541.49, "first_token_was_thinking": false, "wall_s": 2.2745, "eval_count": 256, "eval_duration_ns": 1729941000, "decode_tok_s": 147.982, "prompt_eval_count": 536, "prompt_eval_duration_ns": 80329000, "prefill_tok_s": 6672.559, "load_duration_ns": 456661518, "total_duration_ns": 2271086669, "streamed_chunks": 256, "response_chars": 1033, "thinking_chars": 0, "done_reason": "length", "decode_tok_s_wall": 147.14, "decode_tok_s_wall_gross": 112.551, "num_ctx_option": 131072, "power": { "idle_baseline": { "seconds": 3.0, "samples": 6, "power_w": { "min": 102.95, "median": 102.96, "max": 103.1, "mean": 103.0, "n": 6 }, "utilization_pct": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 6 }, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 6 }, "power_w_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 103.0 }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "power_w_all_cards": 103.0, "ups_whole_box": { "load_pct": null, "realpower_w": null, "realpower_nominal_w": null, "status": null, "note": "the UPS's reading of EVERYTHING it feeds -- the whole box, not the cards. The card figures beside it are board power from nvidia-smi. Neither is the other." } }, "measured_over": "samples where the GPU was busy", "busy_samples": 5, "mean_w": 282.64, "max_w": 331.18, "median_w": 311.62, "mean_w_whole_window": 205.66, "max_w_whole_window": 331.18, "limit_w": 350.0, "over_idle_w": 179.64, "samples": 9, "per_gpu": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 103.0, "median": 207.21, "max": 331.18, "mean": 205.66, "n": 9 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 9 }, "utilization_pct": { "min": 0.0, "median": 91.0, "max": 91.0, "mean": 50.6, "n": 9 }, "temperature_c": { "min": 51.0, "median": 56.0, "max": 57.0, "mean": 54.6, "n": 9 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 48.0, "max": 49.0, "mean": 48.3, "n": 9 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 }, "clocks_sm_mhz": { "min": 1920.0, "median": 2040.0, "max": 2055.0, "mean": 2003.3, "n": 9 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 9 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 9, "busy": { "index": "0", "name": "NVIDIA GeForce RTX 3090 Ti", "power_w": { "min": 207.21, "median": 311.62, "max": 331.18, "mean": 282.64, "n": 5 }, "power_limit_w": 350.0, "memory_used_mib": { "min": 20142.0, "median": 20142.0, "max": 20142.0, "mean": 20142.0, "n": 5 }, "utilization_pct": { "min": 91.0, "median": 91.0, "max": 91.0, "mean": 91.0, "n": 5 }, "temperature_c": { "min": 56.0, "median": 56.0, "max": 57.0, "mean": 56.4, "n": 5 }, "memory_temperature_c": null, "fan_pct": { "min": 48.0, "median": 48.0, "max": 48.0, "mean": 48.0, "n": 5 }, "pcie_width_current": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 5 }, "clocks_sm_mhz": { "min": 2040.0, "median": 2040.0, "max": 2055.0, "mean": 2043.0, "n": 5 }, "clocks_mem_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 5 }, "throttle_sw_power_cap_fraction": 0.0, "throttle_hw_slowdown_fraction": 0.0, "throttle_sw_thermal_fraction": 0.0, "throttle_hw_thermal_fraction": 0.0, "n": 5 }, "busy_samples": 5 } } }, "energy_j_per_1k_tokens": 1909.96, "power_mean_w_all_cards": 282.64, "energy_j_per_1k_tokens_all_cards": 1909.96, "watts_per_tok_s": 1.91, "thermal": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.0, "fan_max_pct": 48.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 282.64, "power_pct_of_limit": 80.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" }, "thermal_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "rule_version": 2, "measured_over": "samples where the GPU was busy (utilisation > 0)", "busy_samples": 5, "temp_max_c": 57.0, "temp_min_c": 56.0, "temp_rise_c": 1.0, "memory_temp_max_c": null, "memory_temp_median_c": null, "memory_temp_readable": false, "pcie_width_min": 16.0, "pcie_width_max": 16.0, "pcie_width_median": 16.0, "fan_mean_pct": 48.0, "fan_max_pct": 48.0, "clock_floor_mhz": 2040.0, "clock_max_mhz": 2055.0, "clock_median_mhz": 2040.0, "mem_clock_floor_mhz": 10251.0, "mem_clock_median_mhz": 10251.0, "mem_clock_max_mhz": 10251.0, "mem_clock_held": true, "memory_bus_width_bits": 384, "peak_mem_bandwidth_gbs_at_median_clock": 984.1, "peak_mem_bandwidth_gbs_at_floor_clock": 984.1, "peak_mem_bandwidth_gbs_at_max_clock": 984.1, "peak_mem_bandwidth_formula": "GB/s = clocks.mem MHz x 1e6 x 2 bits-per-clock x (bus bits / 8) / 1e9; a CEILING at the clock that was observed, never an achieved rate", "clock_dropped": false, "power_limit_w": 350.0, "power_mean_w": 282.64, "power_pct_of_limit": 80.8, "power_pinned_at_limit": false, "sw_power_cap_active_fraction": 0.0, "hw_slowdown_active_fraction": 0.0, "sw_thermal_active_fraction": 0.0, "hw_thermal_active_fraction": 0.0, "at_throttling_temperature": false, "stop_core_temp_c": 83.0, "stop_memory_temp_c": 100.0, "hit_core_temp_stop": false, "hit_memory_temp_stop": false, "throttle_temperature_c": 80.0, "temp_rising_within_run": false, "verdict": "no clock drop during decode" } }, "decode_bandwidth": { "model_tag": "gemma4:26b", "model_density": "moe", "weight_bytes": 17987581215, "peak_bandwidth_gbs_at_observed_clock": 984.1, "decode_gbs": null, "fraction_of_peak_pct": null, "basis": null, "refused_reason": "gemma4:26b is moe: this bench has no receipt for how many bytes a token actually reads, so no bytes-per-second figure is computed. An MoE reads a subset of its weights per token and the subset is not measured here." }, "memory_used_mib_per_card": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": 20142.0 }, "pcie_width_under_load": { "GPU-1809d1e1-1372-f6aa-b687-77ac70675fc4": { "min": 16.0, "median": 16.0, "max": 16.0, "mean": 16.0, "n": 9 } }, "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": 147.691, "median": 147.823, "max": 147.982, "mean": 147.832, "n": 3 }, "ttft_ms": { "min": 538.38, "median": 541.49, "max": 550.9, "mean": 543.59, "n": 3 }, "power_mean_w": { "min": 254.06, "median": 282.64, "max": 284.4, "mean": 273.7, "n": 3 }, "power_mean_w_all_cards": { "min": 254.06, "median": 282.64, "max": 284.4, "mean": 273.7, "n": 3 }, "energy_j_per_1k_tokens_all_cards": { "min": 1720.22, "median": 1909.96, "max": 1923.92, "mean": 1851.37, "n": 3 }, "energy_j_per_1k_tokens": { "min": 1720.22, "median": 1909.96, "max": 1923.92, "mean": 1851.37, "n": 3 }, "mem_clock_median_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "mem_clock_floor_mhz": { "min": 10251.0, "median": 10251.0, "max": 10251.0, "mean": 10251.0, "n": 3 }, "peak_mem_bandwidth_gbs_at_median_clock": { "min": 984.1, "median": 984.1, "max": 984.1, "mean": 984.1, "n": 3 }, "decode_gbs": null, "decode_bandwidth_fraction_of_peak_pct": null, "sw_power_cap_active_fraction": { "min": 0.0, "median": 0.0, "max": 0.0, "mean": 0.0, "n": 3 }, "temp_max_c": { "min": 57.0, "median": 57.0, "max": 57.0, "mean": 57.0, "n": 3 }, "fan_max_pct": { "min": 48.0, "median": 49.0, "max": 52.0, "mean": 49.7, "n": 3 } }, "finished_utc": "2026-09-17T20:11:30Z" } ], "errors": [], "context_length_max": 262144, "quantization": "Q4_K_M", "model_store_size_bytes": 17987581215, "model_store_size_gib": 16.75, "model_density": "moe", "ups_after_arm": { "ups": "pr1500@localhost", "read_utc": "2026-09-17T20:11:30Z", "available": false, "error": "upsc rc=1: Error: Connection failure: Connection refused" }, "largest_context_that_fits": 131072, "levels_requested": [ 4096, 8192, 16384, 32768, 49152, 65536, 98304, 131072 ], "levels_not_reached": [], "ladder_topped_out": "every rung tested fits, up to and including 131,072 -- the ladder ran out of rungs before the card ran out of memory, so the figure reported is the largest window TESTED, not a measured ceiling", "headline": "the biggest context this arm holds for gemma4:26b is 131,072 tokens", "finished_utc": "2026-09-17T20:11:30Z" }