As pytorch uses cudnn under the hood for gpu, what does pytorch use under the hood for cpu?
We have our own implementations for most things just like for cuda. And when possible, we use optimized libraries like cudnn or magma on GPU and mkldnn or mkl/openblas on CPU.
Thank you for your reply. This is helpful.
Is there a commonly preferred default choice out of these options for pytorch operations on CPU?
The default choice in the binaries we provide is mkl for CPU backend.