Does the speed performance have anything to do with the shape of the multiplication matrix? Why is the matrix multiplication of [34,2048]*[2048,5632] much faster than the matrix multiplication of [34,5632]*[5632,2048]? They have the same FLOPs.

[lijianxing123]

No description provided.

[kaleid-liner]

Can you provide me more data? The former one can be slightly faster due to pre computation. However, according to our profiling, the latency of [32, 4096] x [4096, 11008] and [32, 11008] x [11008, 4096] is very close.