A computer I’m working on has 3 GPUs. In the bandwidth test, I found that when
cudaSetDevice(0), the bandwidth is ~6.5GB / s. Meanwhile, on GPU 1/2, it’s ~13GB/s, which is the expected bandwidth for PCIE 3x16.
But GPU 0 is used as the default, so this setup would require me to add
cudaSetDevice() or Python equivalent whenever I want to do something.
Is there any way to select which GPU to be used as GPU0 ?