GPU utilization is too low

when i ran my lstm code on GPU, the GPU utilization is always a few percent like 2%, but the used dedicated memory did increase about 1G.

the question is, was i really using my GPU to accelerate my training process?

i used caffe before, caffe will have a high GPU utilization rate. What’s the difference between torch and caffe when using GPU?

you are using the gpu.

maybe you are either bottlenecked by data loading, or your lstm is so small that launching the gpu compute is slower than actually computing on the gpu.

1 Like