"After core counts, the next battle will be on the interconnect. Low power, scalable, and high performance: process node scaling will mean nothing if the interconnect becomes 90% of the total chip power."
"After core counts, the next battle will be on the interconnect. Low power, scalable, and high performance: process node scaling will mean nothing if the interconnect becomes 90% of the total chip power."
Meaning "highly vectorizable number crunching that you can't practically run on a GPU for some reason" is probably fairly niche. In LINPACK the even the 1950x beat the 7980XE on phoronix: https://www.phoronix.com/scan.php?page=article&item=amd-linu... and I'd have thought LINPACK would really reward AVX...
Maybe it isn't optimized to take advantage of it? I love wasting time playing with number crunchinng microbenchmarks, and when you do that avx-512 seems so unassailably impressive, it's surprising to see this not bourne out in larger benchmarks. (Also worth pointing out you need to tell gcc -mprefer-vector-width=512 on top of march=skylake-avx512, otherwise it'll prefer 256 bit vectors).
Anyway, those benchmarks definitely support compiling and number crunching. Pretty ideal for stuff like monte carlo.