3 ms·
Intel does avx512 at 30+% slower clockspeeds. AMD should be able to do an avx512 at full clockspeeds in double the cycles. For equivalent architectures, you'd
by hajile 8y ago
Intel does avx512 at 30+% slower clockspeeds. AMD should be able to do an avx512 at full clockspeeds in double the cycles.
For equivalent architectures, you'd expect Intel to be faster at workloads that are only avx512. They'd be slower at mixed workloads (the avx unit slows down, but so do all the other ALUs executing in parallel along with the SMT thread).
Most importantly, 512 is 256 more sets of hardware which is a big addition to core size and power usage for what is a very fringe workload. Saving that while getting better performance in common workloads seems like a great tradeoff.
- m0zg 8y agoI'm aware. But it's not just about clock speed per se. AVX512 has some _very_ specialized instructions that are specifically designed to speed up matrix-matrix and matrix-vector multiply. Even at a significantly slower clock those instructions improve things by a lot if you do linear algebra. Most of linalg is done on GPUs nowadays, but not all of it.
- bitL 8y agoAs 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.
- m0zg 8y agoI 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.