I’m working with deep learning using tensorflow, CUDA and CUDNN. Running on Windows 10.
I have previously gotten pretty good GPU utilization while training my CNN models (40%-70%) but now after the big Win 10 10H2 update, the utilization and performance has dropped to almost zero. I have two machines with different cards and CUDA versions (9.0 and 10.1, CUDNN respectively), both working similarly after the update.
Current situation: Memory usage is fine but GPU utilization is something between 4-10% while fitting a model. As a result, training time has increased drastically.
Are there any known issues regarding this and what should I try to fix this?
Thanks beforehand and tell me if this is a wrong board for this issue.