3 ms·
It's possible that AVX512 sort does better on newer architectures, no? Author's machine is a 3 generations old cpu.
by PartiallyTyped 3y ago
It's possible that AVX512 sort does better on newer architectures, no? Author's machine is a 3 generations old cpu.
- pstuart 3y agoIIRC, there were issues with it causing frequency throttling on Intel cpus, whereas AMD's avoid that by "double pumping". It would be very interesting to compare and contrast there. Seems like there's be value in it for all the new ML hotness that's come about. The AMD 7950x seems like it hits the sweet spot for that.
- threadl0cal 3y agoOnly one way to find out -- benchmarks, innit?
- slashdev 3y agoYeah, that would probably change the calculus. There were major frequency (heat) issues on older Intel cpus when running avx512.
- PartiallyTyped 3y agoThere's also the complication that AVX512 was removed in 12xxx and down, and iirc has a softlock in 11xxx.
- celrod 3y agoThe vqsort README says it is a non-issue on that generation CPU (Skylake-X) based on their benchmarks: https://github.com/google/highway/tree/master/hwy/contrib/sort#study-of-avx-512-downclocking https://github.com/google/highway/tree/master/hwy/contrib/so... At least on their low clock speed server CPU (<=3GHz), downclocking was hard to measure compared to clock speed variability they got with std::sort. The 10980xe used for windows benchmarks here normally boosts much higher, so bigger differences could be expected. The author of OP mentioned measuring clock speeds with perf and seeing some difference.