{
 "file": "gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json",
 "model_tag": "mistral-nemo:12b-instruct-2407-q5_K_M",
 "arm": "control-3090",
 "box_class": "gpu-ql3090-24g",
 "quantization": "Q5_K_M",
 "store_gib": 8.13,
 "largest_whole": 131072,
 "topped_out": true,
 "stopped_at": null,
 "errors": [],
 "ups_before": 12.0,
 "ups_after": 20.0,
 "windows": [
  {
   "num_ctx": 4096,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 8647.0
   },
   "runs": 3,
   "decode_median": 78.942,
   "ttft_ms_median": 516.45,
   "board_w_median": 264.11,
   "j_per_1k_median": 3345.62,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 57.0,
   "spread": 0.0009,
   "decode_range": 0.068,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.942
  },
  {
   "num_ctx": 8192,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 470.0
   },
   "runs": 3,
   "decode_median": 78.644,
   "ttft_ms_median": 529.55,
   "board_w_median": 265.29,
   "j_per_1k_median": 3376.14,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 61.0,
   "spread": 0.0017,
   "decode_range": 0.131,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.644
  },
  {
   "num_ctx": 16384,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 696.0
   },
   "runs": 3,
   "decode_median": 78.139,
   "ttft_ms_median": 540.55,
   "board_w_median": 265.72,
   "j_per_1k_median": 3389.59,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0036,
   "decode_range": 0.283,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.139
  },
  {
   "num_ctx": 32768,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1392.0
   },
   "runs": 3,
   "decode_median": 78.158,
   "ttft_ms_median": 516.22,
   "board_w_median": 266.52,
   "j_per_1k_median": 3412.83,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0011,
   "decode_range": 0.085,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.158
  },
  {
   "num_ctx": 49152,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1342.0
   },
   "runs": 3,
   "decode_median": 78.267,
   "ttft_ms_median": 516.34,
   "board_w_median": 266.32,
   "j_per_1k_median": 3402.7,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 63.0,
   "spread": 0.0022,
   "decode_range": 0.171,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.267
  },
  {
   "num_ctx": 65536,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1440.0
   },
   "runs": 3,
   "decode_median": 78.016,
   "ttft_ms_median": 490.13,
   "board_w_median": 265.3,
   "j_per_1k_median": 3400.58,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0038,
   "decode_range": 0.299,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.016
  },
  {
   "num_ctx": 98304,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 2880.0
   },
   "runs": 3,
   "decode_median": 78.148,
   "ttft_ms_median": 540.53,
   "board_w_median": 259.14,
   "j_per_1k_median": 3310.76,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 63.0,
   "spread": 0.002,
   "decode_range": 0.157,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.148
  },
  {
   "num_ctx": 131072,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 2880.0
   },
   "runs": 3,
   "decode_median": 78.223,
   "ttft_ms_median": 545.37,
   "board_w_median": 257.51,
   "j_per_1k_median": 3295.74,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 63.0,
   "spread": 0.0023,
   "decode_range": 0.183,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 78.223
  }
 ],
 "resident_named_card_growth_mib": 8647.0,
 "resident": true,
 "passing_windows": [
  4096,
  8192,
  16384,
  32768,
  49152,
  65536,
  98304,
  131072
 ]
}