Implement better metric for automatically choosing the best GPU
denisalevi opened this issue · 0 comments
denisalevi commented
Currently, we choose the GPU based on compute capability, which is basically picking new GPUs over potentially faster ones (see #160).
I just saw a findCudaDevice function in the CUDA samples, which chooses the GPU with highest performance = #multiprocessors * #SM-per-multiprocessor * clock-rate. Seems reasonable.
Found it in NVIDIA_CUDA-11.2_Samples/common/inc/helper_cuda.h