I am doing lots of Fourier Transforms of large datasets using tensorflow-gpu.
On the Windows 10 task manager, I see that the GPU "cuda" engine is only at 5% usage. The GPU dedicated memory is at 4.8/6 GB (but I also have questions about this).
Because of the low usage of the 'cuda' engine, I am assuming there are other bottlenecks. The obvious one is delays in transferring from CPU to GPU memory.
Are there any diagnostics in python or Windows that can show how often CPU memory is accessed?
So far I am only trying to use tf.device on all my tensorflow objects to see if they are indeed pinned to the GPU.