[
  {
    "model": "pe-av-small-16-frame",
    "model_id": "facebook/pe-av-small-16-frame",
    "revision": "9f888eea95c83622212bb742e91bf01d3b46fe96",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 846991308,
    "model_load_seconds": 1.405979960982222,
    "text_bank_seconds": {
      "imagenet1000": 2.8613254200317897,
      "kinetics400": 0.7043806570000015,
      "audioset527": 0.9188710229936987,
      "visual_motion5": 0.16409541602479294
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": false,
    "peak_allocated_bytes": 4196955648,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 1.462812562880572,
        "encoder_mean_seconds": 0.02925625125761144,
        "encoder_median_seconds": 0.029262166004627943,
        "encoder_p95_seconds": 0.029332908650394528,
        "standalone_total_mean_seconds": 1.3892483090981842,
        "million_encoder_gpu_hours": 8.126736460447622,
        "million_standalone_gpu_hours": 385.90230808282894
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 4.354261539119761,
        "encoder_mean_seconds": 0.08708523078239523,
        "encoder_median_seconds": 0.06838601452182047,
        "encoder_p95_seconds": 0.17645975194172916,
        "standalone_total_mean_seconds": 0.14356576248421335,
        "million_encoder_gpu_hours": 24.190341883998673,
        "million_standalone_gpu_hours": 39.879378467837036
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 4.314490076969378,
        "encoder_mean_seconds": 0.08628980153938755,
        "encoder_median_seconds": 0.0873057214776054,
        "encoder_p95_seconds": 0.10171568599471356,
        "standalone_total_mean_seconds": 1.5027623910817782,
        "million_encoder_gpu_hours": 23.969389316496542,
        "million_standalone_gpu_hours": 417.4339975227162
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  },
  {
    "model": "pe-av-base-16-frame",
    "model_id": "facebook/pe-av-base-16-frame",
    "revision": "3fd870e60a1099fb99367b240d41926e183d7112",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 1033701580,
    "model_load_seconds": 14.907948495005257,
    "text_bank_seconds": {
      "imagenet1000": 2.8377642179839313,
      "kinetics400": 0.6878930119564757,
      "audioset527": 0.9234632209991105,
      "visual_motion5": 0.052366399962920696
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": false,
    "peak_allocated_bytes": 4539169792,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 1.4475287970271893,
        "encoder_mean_seconds": 0.028950575940543784,
        "encoder_median_seconds": 0.028947740036528558,
        "encoder_p95_seconds": 0.029030600254191085,
        "standalone_total_mean_seconds": 1.3547041056049056,
        "million_encoder_gpu_hours": 8.041826650151052,
        "million_standalone_gpu_hours": 376.3066960013627
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 3.590409246913623,
        "encoder_mean_seconds": 0.07180818493827246,
        "encoder_median_seconds": 0.07003838248783723,
        "encoder_p95_seconds": 0.10310203270346392,
        "standalone_total_mean_seconds": 0.12713899948983454,
        "million_encoder_gpu_hours": 19.946718038409017,
        "million_standalone_gpu_hours": 35.31638874717626
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 4.475306674023159,
        "encoder_mean_seconds": 0.08950613348046317,
        "encoder_median_seconds": 0.08986335751251318,
        "encoder_p95_seconds": 0.10575635436398442,
        "standalone_total_mean_seconds": 1.4705904776963872,
        "million_encoder_gpu_hours": 24.862814855684217,
        "million_standalone_gpu_hours": 408.4973549156631
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  },
  {
    "model": "pe-av-large-16-frame",
    "model_id": "facebook/pe-av-large-16-frame",
    "revision": "446d823f089b1301a0ff37175ab283b6927db757",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 2234700748,
    "model_load_seconds": 23.750902826024685,
    "text_bank_seconds": {
      "imagenet1000": 1.5821360899717547,
      "kinetics400": 0.6589782600058243,
      "audioset527": 0.8984351100516506,
      "visual_motion5": 0.017756743007339537
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": false,
    "peak_allocated_bytes": 6944590848,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 1.4823009388055652,
        "encoder_mean_seconds": 0.029646018776111303,
        "encoder_median_seconds": 0.029642128007253632,
        "encoder_p95_seconds": 0.029727862108848056,
        "standalone_total_mean_seconds": 1.3969909888354595,
        "million_encoder_gpu_hours": 8.235005215586474,
        "million_standalone_gpu_hours": 388.05305245429435
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 4.384937118098605,
        "encoder_mean_seconds": 0.08769874236197212,
        "encoder_median_seconds": 0.0862295079859905,
        "encoder_p95_seconds": 0.11656579277187118,
        "standalone_total_mean_seconds": 0.14469541802769526,
        "million_encoder_gpu_hours": 24.360761767214473,
        "million_standalone_gpu_hours": 40.193171674359796
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 5.1698952646693215,
        "encoder_mean_seconds": 0.10339790529338642,
        "encoder_median_seconds": 0.10231557849328965,
        "encoder_p95_seconds": 0.1187008213775698,
        "standalone_total_mean_seconds": 1.527739551018458,
        "million_encoder_gpu_hours": 28.72164035927401,
        "million_standalone_gpu_hours": 424.37209750512716
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  },
  {
    "model": "pe-av-small",
    "model_id": "facebook/pe-av-small",
    "revision": "dd050762bb9704ae9cd996ca45532a98f81d817e",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 846991308,
    "model_load_seconds": 13.106045840017032,
    "text_bank_seconds": {
      "imagenet1000": 1.7188300080015324,
      "kinetics400": 0.6564783719950356,
      "audioset527": 0.8942986779729836,
      "visual_motion5": 0.017920163983944803
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": true,
    "peak_allocated_bytes": 8376205824,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 52.481551192875486,
        "encoder_mean_seconds": 1.0496310238575097,
        "encoder_median_seconds": 0.9454118104767986,
        "encoder_p95_seconds": 1.6290604809037177,
        "standalone_total_mean_seconds": 2.469762842853088,
        "million_encoder_gpu_hours": 291.5641732937527,
        "million_standalone_gpu_hours": 686.0452341258577
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 2.5585094449925236,
        "encoder_mean_seconds": 0.05117018889985047,
        "encoder_median_seconds": 0.0505299994838424,
        "encoder_p95_seconds": 0.06720614168443717,
        "standalone_total_mean_seconds": 0.10795936825801618,
        "million_encoder_gpu_hours": 14.213941361069576,
        "million_standalone_gpu_hours": 29.988713405004496
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 53.279276007262524,
        "encoder_mean_seconds": 1.0655855201452504,
        "encoder_median_seconds": 0.9910256644943729,
        "encoder_p95_seconds": 1.6683923832169967,
        "standalone_total_mean_seconds": 2.5425065184989943,
        "million_encoder_gpu_hours": 295.99597781812514,
        "million_standalone_gpu_hours": 706.2518106941651
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  },
  {
    "model": "pe-av-base",
    "model_id": "facebook/pe-av-base",
    "revision": "1c4c329b78d80a18d860ebfb05daebff4bc44518",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 1033701580,
    "model_load_seconds": 14.913788648031186,
    "text_bank_seconds": {
      "imagenet1000": 1.677044419979211,
      "kinetics400": 0.6609332300140522,
      "audioset527": 0.8498661980265751,
      "visual_motion5": 0.017385102983098477
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": true,
    "peak_allocated_bytes": 8719813632,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 53.17707831325242,
        "encoder_mean_seconds": 1.0635415662650485,
        "encoder_median_seconds": 0.9242103355063591,
        "encoder_p95_seconds": 1.6785006219404748,
        "standalone_total_mean_seconds": 2.460226884263102,
        "million_encoder_gpu_hours": 295.42821285140235,
        "million_standalone_gpu_hours": 683.3963567397506
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 2.7695929193287157,
        "encoder_mean_seconds": 0.055391858386574316,
        "encoder_median_seconds": 0.05549991253064945,
        "encoder_p95_seconds": 0.07096275268995669,
        "standalone_total_mean_seconds": 0.11117335592862218,
        "million_encoder_gpu_hours": 15.386627329603975,
        "million_standalone_gpu_hours": 30.88148775795061
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 53.15067586291116,
        "encoder_mean_seconds": 1.0630135172582231,
        "encoder_median_seconds": 0.9619682100310456,
        "encoder_p95_seconds": 1.6608091613627038,
        "standalone_total_mean_seconds": 2.5154803327983246,
        "million_encoder_gpu_hours": 295.28153257172863,
        "million_standalone_gpu_hours": 698.7445368884235
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  },
  {
    "model": "pe-av-large",
    "model_id": "facebook/pe-av-large",
    "revision": "0d24878d4107d64bef49e53602fc34ce6f94f6d8",
    "gpu": "NVIDIA GH200 120GB",
    "gpu_total_memory_bytes": 102005473280,
    "dtype": "bfloat16",
    "clip_batch_size": 1,
    "vision_frame_chunk_size": 16,
    "model_parameter_count": 2234700748,
    "model_load_seconds": 28.547460709989537,
    "text_bank_seconds": {
      "imagenet1000": 1.6945692419540137,
      "kinetics400": 0.6769166039885022,
      "audioset527": 0.853914042992983,
      "visual_motion5": 0.016733628988731652
    },
    "warmup_each_mode": 1,
    "clip_limit_seconds": 30,
    "native_frames": true,
    "peak_allocated_bytes": 11120736768,
    "requested_records": 50,
    "successful_records": 50,
    "error_records": 0,
    "modes": {
      "video": {
        "n": 50,
        "encoder_total_seconds": 51.222722226870246,
        "encoder_mean_seconds": 1.0244544445374049,
        "encoder_median_seconds": 0.9058442884706892,
        "encoder_p95_seconds": 1.6160722666449145,
        "standalone_total_mean_seconds": 2.4220211274246686,
        "million_encoder_gpu_hours": 284.570679038168,
        "million_standalone_gpu_hours": 672.7836465068524
      },
      "audio": {
        "n": 50,
        "encoder_total_seconds": 4.39572861790657,
        "encoder_mean_seconds": 0.08791457235813141,
        "encoder_median_seconds": 0.08372071749181487,
        "encoder_p95_seconds": 0.10954021530633326,
        "standalone_total_mean_seconds": 0.14385590896126815,
        "million_encoder_gpu_hours": 24.420714543925392,
        "million_standalone_gpu_hours": 39.959974711463374
      },
      "audio_video": {
        "n": 50,
        "encoder_total_seconds": 53.5818614558666,
        "encoder_mean_seconds": 1.071637229117332,
        "encoder_median_seconds": 0.9727511060191318,
        "encoder_p95_seconds": 1.6718731961096638,
        "standalone_total_mean_seconds": 2.5251452486077324,
        "million_encoder_gpu_hours": 297.6770080881478,
        "million_standalone_gpu_hours": 701.4292357243701
      }
    },
    "normalization": "All saved clip embeddings L2-normalized. Text: normalize each template embedding, mean templates, renormalize.",
    "preprocessing": "First<=30seconds; native decoded video frames (16uniformframes for -16-frame variants); PyAV bilinear square336RGB resize; values scaled to[-1,1]; mono48kHz audio; reflect-pad to1920sample multiple.",
    "extrapolation": "Linear estimate from observed mean seconds per successfully encoded clip, batch1 and this hardware/precision. Excludes downloads, model/text startup, scheduling, storage scaling and failures. No claim of production throughput.",
    "score_precision": "float32 cosine and softmax; encoder inference BF16"
  }
]