GPU h.264 decoding speed


I’m using Gstreamer and the nvcodec element to decode rtp h.264 network video. I aim for shortest possible latency. When measuring latency it seems that decoding a frame takes on average 40-60 milliseconds with 720P@25fps on a Quadro M2000M card. This measurement comes from the built in Gstreamer tracing.

Averaging 40-60 milliseconds to decode a frame seems to be a very long time as I see it. Am I expecting too much when it comes to GPU decoding?

I don’t know if it helps you but when I max out a T4 decoding H264 720P resolution I get 3020 FPS meaning a decode speed of ~0.3ms per frame. I write my own stuff using C++ / CUDA and I make sure I stay in GPU memory in my pipelines.

So yes 40-60ms is slow but better ask the Gstreamer people what to expect maybe?


Thanks Anders. I’m already bugging the GStreamer people. :-)