Re:AMD Ryzen 3rd Gen 'Matisse' Coming Mid 2019: Eight Core Z...
by bitL on Wednesday January 09, 2019 @06:41PM
(#18870180)
As a consequence, TR/Ryzen is very slow in "classical" ML in Python (I can tell), because MKL/AVX2 seems to be way slower. This could be fixed with Zen 2; AVX512 then will be likely at 2/3 of Intel's performance without slowing down the rest of the system during computations.
[ Parent
| on HN ]
Re:AMD Ryzen 3rd Gen 'Matisse' Coming Mid 2019: Eight Core Z...
by m0zg on Wednesday January 09, 2019 @07:52PM
(#18870579)
I agree with this assessment. Frankly, GPUs are so much faster for dense linalg than even the highest end Intel (or AMD) chips that anything that can be done on GPUs should be done there. But that requires one to know how to work with them, if the primitives don't already exist, and most people are oblivious to the finer points.