{"ran_at":"2026-11-02T19:20:50.775600Z","version":"0.1.0","onnxruntime":"1.29.0","ep":"tensorrt","device":0,"tflops":7.316,"model_load_ms":{"yolox-s@1.0.0":15502.0},"model_ms":{"yolox-s@1.0.0":{"fp32":4.916,"fp16":3.421,"input":"640x640"}},"detector_fps":{"b1":151.5,"b4":null,"b8":null,"fp16_b1":199.9},"decode_fps":{"cpu":2476.2,"cpu_per_session":2476.2,"cpu_sessions":1,"hw":1950.9,"hw_per_session":1184.7,"hw_sessions":2,"hw_decoder":"cuda","fullpic_cpu":2789.8,"fullpic_hw":1729.5},"capacity":{"streams_at_fps":{"fps":10.0,"streams":15},"concurrent_models":3,"basis":{"detector_fps":151.47465919836506,"decode_fps":1950.9200336107904,"source_fps":12.0,"shared_cpu":false,"batching":false,"decode":"hw","at_least":false},"by_type":{"det5":30,"crowd2":75,"anpr":null,"face":null,"search":null},"decodes":144,"precision":"fp32","basis_by_type":{"models":{"det":"yolox-s@1.0.0"},"ms":{"det":6.60176431683164},"fps":{"det5":5.0,"crowd2":2.0,"anpr":5.0,"face":5.0,"search":5.0},"plates_per_s":2.0,"faces_per_s":5.0,"crops_per_s":2.0,"decode_fps":1950.9200336107904,"fullpic_fps":1729.530807698672,"source_fps":12.0,"shared_cpu":false,"at_least":[]}},"clip":{"width":768,"height":432,"fps":12.0},"notes":{"capacity.scope":"the detector and video decoding only: plate reading (OCR) and embeddings (CLIP) cost more per stream and are not included","detector_fps.b4":"the model's batch dimension is fixed at 1","detector_fps.b8":"the model's batch dimension is fixed at 1","capacity.by_type":"streams per type at the catalogue rates (fps: det5 5, crowd2 2, anpr 5, face 5, search 5) and 2 plates/s, 5 faces/s, 2 crops/s: cost_s = fps\u00b7det_s (det5, crowd2); fps\u00b7det_s + plates/s\u00b7(scout_s + 2\u00b7lexa_s) (anpr); fps\u00b7visage_s + faces/s\u00b7persona_s (face); fps\u00b7det_s + crops/s\u00b7recall_s (search); det_s from the detector loop, the others one inference at batch 1 (scout_s at 640x640, ANPR's vehicle crops); streams = floor(min(1/cost_s, V/R)) (separate engines), V the letterbox decode (the full-picture decode for anpr), R the clip's rate; no headroom applied; null = a model it needs is not installed or a figure is missing","capacity.decodes":"the slower of the letterbox and full-picture decodes (full picture) at the clip's rate"}}