{ "profile": "cuda resident depth, CPU decode for CPU consumers", "basis": "code audit; raw host byte counters cover explicit application frame buffers, not an Nsight trace", "removed": [ "NVDEC GPU decoded image -> CPU image (software decode now)", "DA3 GPU depth -> NumPy CPU depth -> CUDA depth", "CUDA gray -> CPU gray pipe -> NVENC GPU input", "Unused DA3 confidence/sky/camera output conversions and processed image visualization" ], "remaining": [ "CPU decoded uint8 RGB -> DINO GPU input, FP32 conversion after transfer", "DA3 CPU official preprocessed tensor -> GPU model input", "Four percentile endpoint scalars and scalar validation results -> CPU", "NVENC compressed bitstream -> CPU/network", "CPU ffprobe frame/timestamp validation" ], "not_removed": [ "Source CPU decoding occurs in both filter and depth stages; no shared RGB cache" ], "no_npz": true }