We are currently experiencing issues when decoding video using the Video Codec and SDK running on the NVIDIA Deep Learning AMI (https://aws.amazon.com/marketplace/pp/B076K31M1S?ref=cns_srchrow) using g4dn.2xlarge instances. When attempting to run the samples with a 2 second AVC 720p video it takes roughly 30 minutes to decode the video. It seems to correct itself after the very first run, but on reboot the issue pops up again. This issue only happens on the latest Tesla driver 440.33.01 and earlier driver versions didn’t have this issue.
I have attached the sample code with a Dockerfile that builds the NVIDIA Video Codec SDK sample apps and then a run.sh script that runs the test. You will see from the test that the attached 2 second video takes roughly 30 minutes to decode. It seems to work fine after the first run, but taking this long for the first decode does not work for our application needs.
Code to create test video and run the samples can be found here: https://gist.github.com/camdenfullmer/c0e6a83d5f14330ae7bbf3f465967d65
Output when running test:
root@ip-172-31-13-241:/home/ubuntu# time ./run.sh
GPU in use: Tesla T4
Decode with demuxing.
[INFO ][04:50:07] Media format: QuickTime / MOV (mov,mp4,m4a,3gp,3g2,mj2)
Session Initialization Time: 449561 ms
Total frame decoded: 60
Saved in file out.native in nv12 format
Session Deinitialization Time: 536737 ms