When comparing video memory usage from the exact same config (hard + soft) and simply changing the GPU, it seems that more recent generations consistently use drastically more on my use cases.
The tests were run on a custom CentOS 7.9, swapping out between Quadro P4000, Quadro RTX4000, and Quadro RTX-A4000.
Drivers 450.x, 470.x and 510.x were tried but this did not seem to modify the observed results.
The programs that were tried run a similar task of video ingest, processing and encoding.
These are the nvidia-smi values for these programs:
- OpenGL + CUDA + NVEnc => | P4000: 810 MiB | RTX4000: 893 MiB | RTX-A4000: 1904 MiB |
- CUDA + NVEnc => | P4000: 174 MiB | RTX4000: 194 MiB | RTX-A4000: 258 MiB |
- CUDA + NVJpeg => | P4000: 199 MiB | RTX4000: 221 MiB | RTX-A4000: 285 MiB |
There is no specific conditional code in these programs that would change the behavior depending on device architecture.
So I have 2 questions about this:
- Is this a known issue that others have encountered?
- If it is not, what kind of mistake could cause this phenomenon?
Thanks in advance,