These functions (cudnnFindConvolutionForwardAlgorithm,cudnnFindConvolutionBackwardDataAlgorithm, …) run all (or some of) supported convolutions algorithms and return their numbers in execution time order. I have noticed that the fastest algorithm returned by one of these functions for the same convolution may change between runs.
As I understand CUDA kernels run on a GPU exclusivley unless MultiProcessService is used. Now, I am just wondering if one process using GPU may affect results of cudnnFindConvolution*Algorithm functions in other processes sharing same GPU through MPS.