I am using 3D Conv in my model.
I notice that when I scale out my model to extremely large.
The kernel volta_scudnn_128x64_stridedB_splitK_medium_nn_v1 becomes inefficient.
Could anyone explain me the function of the kernel?
If you are using cudnn 7,6 then 3d support is limited. You can use cudnn v8 which has better 3d support and also backend APIs that allows you to select fastest kernels for your input size.
Please refer to the below links for more details.