Or are computations offloaded to CPU? I noticed that my CPU utilization jumped from 10-20% to 100% after toggling torch.backends.cudnn.enabled from true to false.
Yes, PyTorch will use native CUDA kernels if cuDNN is disabled.
Thanks so much for the fast reply!! Do you have any insight into why setting torch.backends.cudnn.enabled to false would drastically increase CPU utlilization like this?
No, but you could profile the workload to check where the increased CPU workload is coming from.