NVENC fails with CUDA error 999

Hi, all.

I’m transcoding videos with NvTranscoder sample program on Quadro P5000, it seems that when it come to more and more processes, the program will fail with this error:

cuCtxCreate (&cudaCtx, CU_CTX_SCHED_AUTO, device) has returned CUDA error 999

While the memory usage has not reaches 100%, does this error mean it reaches the performance neck? Does Video Codec SDK have an defined error for this?

I tried multiple versions of nvidia drivers, it seems the newer version has better performance.



cuCtxCreate is a CUDA API and hence the error code returned is a CUDA error enumeration which is actually CUDA_ERROR_UNKNOWN. Hence, there is no defined error code for this from Video Codec SDK.

I have a couple of questions to understand the problem better.

  1. What is the meaning of “more and more processes”? What does each process do? What is your application trying to do? Encode? Decode? Transcode? Or something else?

  2. When you get this error, what is the memory utilization? How are you measuring it?

  3. When you say the newer driver version has better performance what exactly do you mean? Is it in terms of memory?

Thank you for the help

Hi, thank you for getting back to me.

  1. I use this card for video transcoding, each process transcode one video. The tests use NvTranscoder sample code.

  2. The memory utilization for 64 processes of 480P video is about 45%. I use nvidia-smi to measure. And the 65th process will get this error. When it comes to 1080P video, the processes count can be about 22, and the memory utilization will be near 100%, so no more process to be created to get this error.

  3. No it’s not about the memory, the older driver version, the lower processes can be created when it comes to get this error.