Exactly this. A handful of good AMD engineers should be able to squeeze everything out of their own hardware by tuning fast convolution, gemm and other ML relat

Exactly this. A handful of good AMD engineers should be able to squeeze everything out of their own hardware by tuning fast convolution, gemm and other ML related kernels optimised for Vega. That happens once, not for every framework. AMD's management just doesn't seem to be that interested in ML.