4 ms·
I use some engineering software that uses Intel Math Kernel Libraries. It is 30% faster on an Intel CPU, versus AMD with equivalent speed/cores.
by madengr 6y ago
I use some engineering software that uses Intel Math Kernel Libraries. It is 30% faster on an Intel CPU, versus AMD with equivalent speed/cores.
- jiggawatts 6y agoIs it a 30% perf/$ advantage though? In most comparisons I've seen Intel lost horrendously even with AVX-512 code compared to EPYC 2 when price was factored in. Doubly so if the cost of electricity and cooling is included, because the watts/core for AMD is much lower than anything Intel has.
- threeseed 6y agoYou can see my 10980xe Geekbench 4/5 results here: https://browser.geekbench.com/user/threeseed https://browser.geekbench.com/user/threeseed Compare against 3950x and 3960x and you will see it's closer to the 3960x for quite a bit cheaper and with 6 less cores. And that's not entirely AVX-512 code. Also power consumption is a little overblown since all CPUs scale their frequencies so even though mine runs at 5GHz it will spend most of the day below 2GHz or so.
- LargoLasskhyfv 6y agoAre you aware of the workaround for the mkl to not fall back to SSE on AMD? [1] https://old.reddit.com/r/MachineLearning/comments/f2pbvz/discussion_workaround_for_mkl_on_amd/ https://old.reddit.com/r/MachineLearning/comments/f2pbvz/dis...
- madengr 6y agoNo, thank you! I am going to try this.
- gameswithgo 6y agothat is because intel won’t fix a bug that doesn’t let the amd cpus use avx
- FartyMcFarter 6y agoIt's not a bug, there's a well documented history of Intel artificially crippling performance on non-Intel chips. They've done this with their compiler (and got sued for it), libraries, and they've even introduced code into glibc to do that.