Intel Releases x86-simd-sort 2.0 With Faster AVX-512 Sorting, New Algorithms

[phoronix]

Phoronix: Intel Releases x86-simd-sort 2.0 With Faster AVX-512 Sorting, New Algorithms

Earlier this year Intel software engineers published a blazing fast AVX-512 sorting library that was initially picked up by Numpy where it netted them 10~17x faster sorts. Today marks the release of x86-simd-sort 2.0 with even more AVX-512 features in place and additional sorting algorithms added...

https://www.phoronix.com/news/Intel-x86-simd-sort-2.0

[rene]

Cool, that will run so fast on Ryzen 7000 ;-) https://www.youtube.com/watch?v=iBiofQvJYFw

[quaz0r]

Is there any real reason not to just drop something like this into the standard library

[bumblebritches57]

Why aren’t they contributing this to various libc’s?

hell, even contribute it to LLVM’s overlay libc.

or at least make it public domain so libc authors can use it if they want to, but the last thing anyone needs is Intel releasing yet another platform specific library that people have to opt into.

[spiral_23]

my tiger lake - likes it!

[fitzie]

Why aren’t they contributing this to various libc’s?

hell, even contribute it to LLVM’s overlay libc.

or at least make it public domain so libc authors can use it if they want to, but the last thing anyone needs is Intel releasing yet another platform specific library that people have to opt into.

they are working on getting it into numpy, which seems to be an audience that would benefit from the improvement. doesn't seem like it is a designed to be a drop in replacement to std::sort.

[vegabook]

my tiger lake - likes it!

Lol me too.

Picked up a nice little Lenovo Yoga Slim that had a US keyboard that nobody wanted here in the UK, for the paltry sum of 400 pounds (16/512/Tiger Lake 11th Gen i5-1135G7), which is conspicuous by the absence of its AVX512 being fused off as per gen 12 onwards. Recent Numpy builds love it.

This was launched at around the time that AMD's Zen was ripping into Intel market share so they panicked and left a bit of Xeon-class goodness on the die. Reminds me of when Radeon VII desperately did not fully gimp FP64 in a rearguard action to differentiate from the 20 series Nvidia onslaught.

[coder]

Is there any real reason not to just drop something like this into the standard library

It would lead to performance regressions on Skylake SP and Cascade Lake SP Xeon CPUs (not sure about Ice Lake). They clock-throttle even at very small amounts of AVX-512 usage, which means that unless your app is AVX-512 dominant, using just a little AVX-512 code is worse than none at all.

At some point, I could agree that we care about optimizing for newer CPUs more than coddling old ones. Still seems a bit too recent, as most Cascade Lake Xeons are probably still in service.

BTW, I'm also interested in efficiency. Would be great to get some CPU "package power" or "core power" stats, so we could track the perf/W and not just performance in a vacuum.

[antoinec]

Intel is releasing software ... OH, wait a minute.

Next time, Microsoft will release hardware CPU patches for Windows.

[spiral_23]

Intel is releasing software ... OH, wait a minute.

Next time, Microsoft will release hardware CPU patches for Windows.

and apple will release no more zero day security holes!