I have got a server PC with three GPUs, RTX A6000. I’m using the three GPUs simultaneously with CuPy. The CUDA version I’ve is 12.1.
I can use the three GPUs at the same time. The problem I’m having is that when I open nvidia-smi to check the GPU’s memory usage, GPU #1 uses almost double the memory of GPU #0. The same happens with GPU #2. GPU #2 uses almost double the memory of GPU #0
As I’m creating three independent but equal CuPy arrays (same size and type) on three GPUs, I expected that the occupied memory on each GPU would be the same.
Has anyone faced this issue? Is this normal behaviour?