I’ve got a few questions that maybe you or other blog members can answer.
First Question relates to use of multiple GPUs within a server for video transcoding.
We’ve been using 16 GB P100s to transcode MPEG2 and H264 to HEVC 720p, but have not been able to get 4 GPUs to work simultaneously in one SuperMicro SYS-1029GQ-TRT server, even though all server I/O, CPU, Internal Bus thruput and Memory resources are in the 15-20% range.
At most we’ve been able to transcode about 22 channels (`660 FPS). When we try to add transcodes on a second or third GPU, the total number of channels does not increase. Do any ideas come up as to what may be causing this limitation?
Second Question is:
Would 32 GB Tesla V100s, or Quadro GV100, or Turing TU104s be able to increase the total number of simultaneous live video transcodes versus the Tesla P100s?
Third Question is:
Do you know when the new Video Codec SDK, which exposes new encoder improvements and features of Turing will be released?
I noticed today that FFmpeg source was updated so that it requires the driver to report NVENC API version 8.2.
However, even the latest drivers only report in the FFmpeg log that they have loaded NVEncodeAPI 8.1. Is that correct - if the 8.2 SDK is out, shouldn’t the latest drivers (released just 3 days ago) be reporting 8.2 too?
@oviano: If you checkout the 8.2.15.5 tag, you will see that inside nvEncodeAPI.h the version is still 8.1 . So the issue still exists, and probably is in the driver that still reports 8.1
Hi, I also faced the same problem about 8.2 today but just a few hours ago someone updated the nv-codedc-headers which fixed the issue. All I needed to do was to delete the ffnvcodec-git and ffmpeg-git folders as well as the related files in the local64 folder in the media-autobuild_suite and rebuild. Hope this helps!