When using the CPU decoding ( * ffplay -vcodec h264 1000_raw_data.h264') I noticed video is OK but there are a lot of “Invalid NAL Units” Errors - it is possible that the h264 Parser that is used by the NVIDIA SDK is not handling the Invalid NAL unit properly and try to decode them which lead to a lot of errors. I tried to use the input you provided to the h264 decode sample of the SDK and got a lot of “CONCEALED ERRORS” the resulted video was not “green” as when using ' * ffplay -vcodec h264_cuvid 1000_raw_data.h264' - I did get a few normal frames every few seconds but the video as a whole was not playble (lots of artifacts in the video) The SDK pass the input data through it own h264 parser to identify the h264 nal units and other h264 parameters. The FFMPEG H264 GPU decoding is implemented based on the NVIDIA Video SDK. I will provide my sample data here One Drive

ffmpeg -vcodec h264_cuvid -i 1000_raw_data.h264 output.mp4.

FFMPEG/libavcodec: 4.4.2 (installed from source, following Official Guide and Pytorch Guide.FFMPEG/libavcodec: 4.2.7 (installed using apt).So I think that maybe GPU H264 Codec is bugged, or there is some kind of H264 data format that GPU implementation cannot decode? I also ran a GPU H265 Decoder, and it can work well all day. I tried to use both libavcodec and ffmpeg application on two different servers and GPUs, and they seemed to perform similarly, failed to decode the same data, worked normally for the other. It didn’t always fail through, maybe approx 50% of input data.

I found that codec h264_cuvid failed to decode on some data, while CPU implementation h264 / libx264 worked normally. My implementation is based on FFmpeg/decode_video.c at master

Ffmpeg vcodec h264