Multiply add batches of matrices of shape [n, 4, 4], [n, 4, 4] and [n,4,4] directly on tensor cores about cutlass

[hwu36]

The link Matt posted is for wmma. As to native mma, you can check https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#warp-level-matrix-instructions-for-mma

[guidocalvano]

That link is to the most awesome paper on neural nets I've read by the way. I highly recommend it. It's a shame they take a day to train those nets on 8 graphics cards. Imagine how computer vision would change if they only took a day to train ;)

[mnicely]

At this time, tensor core sizes are limited to the set provided in documentation
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#wmma-type-sizes