Thanks for digging that up. Now, being faster than stock MKL based pytorch or tf as well as being more *cost efficient* I can believe.