VPOPCNTDQ and VBMI2

[xr_a_y]

Did someone already tried VPOPCNTDQ and/or VBMI2 ? what is the expected performance improvment ?

[mar]

Did someone already tried VPOPCNTDQ and/or VBMI2 ? what is the expected performance improvment ?

no idea, but I assume those are vector instructions
for example, arm64 doesn't have popcnt for gpr registers, so you have to load into a vfp register, execute cnt on 8-bit vector, then accumulate per-byte results and convert back to a general purpose register.

doesn't seem to me like this should be a win at all, assuming you want to do popcnt on 64-bit bitboards

[Joost Buijs]

These AVX-512 instructions are only supported by a few Intel architectures like the new Rocket Lake i9-11900K, nowadays everybody seems to buy AMD so there won't be many people with access to one of these processors.

[Ras]

These AVX-512 instructions are only supported by a few Intel architectures

And if you actually use these instructions, the CPU will throttle down so much that it will be a loss in anything but crafted benchmarks for dubious marketing - which is what these instructions are actually for.

[Raphexon]

These AVX-512 instructions are only supported by a few Intel architectures

And if you actually use these instructions, the CPU will throttle down so much that it will be a loss in anything but crafted benchmarks for dubious marketing - which is what these instructions are actually for.

AVX512 is actually a huge speed up if a majority of the instructions are avx512.
In mixed loads the throttle can indeed cause slow downs.

I think vnni256 is fastest for SF, followed by vnni512. But it also depends on the CPU, some throttle less than others.

[Joost Buijs]

These AVX-512 instructions are only supported by a few Intel architectures

And if you actually use these instructions, the CPU will throttle down so much that it will be a loss in anything but crafted benchmarks for dubious marketing - which is what these instructions are actually for.

AVX512 is actually a huge speed up if a majority of the instructions are avx512.
In mixed loads the throttle can indeed cause slow downs.

I think vnni256 is fastest for SF, followed by vnni512. But it also depends on the CPU, some throttle less than others.

Indeed it gives some gain, I never tried with Stockfish but with my own engine the gain with AVX-512 is 20 to 25% in comparison with AVX2. This is on my i9-10980XE. When properly cooled it doesn't throttle much with AVX-512 and still runs at 3800 MHz. all core.

The 8 core i9-11900K is even better in this respect but uses a shitload of power (300W), and this is a lot for just 8 cores. An acquaintance of mine has one, it is extremely fast with AVX2 and AVX-512.

[Modern Times]

The 8 core i9-11900K is even better in this respect but uses a shitload of power (300W), and this is a lot for just 8 cores. An acquaintance of mine has one, it is extremely fast with AVX2 and AVX-512.

And yet, apart from the AVX-512 support, (or even despite the lack of AVX-512 support) you're probably better off with the 10-core 10900K.

[Joost Buijs]

The 8 core i9-11900K is even better in this respect but uses a shitload of power (300W), and this is a lot for just 8 cores. An acquaintance of mine has one, it is extremely fast with AVX2 and AVX-512.

And yet, apart from the AVX-512 support, (or even despite the lack of AVX-512 support) you're probably better off with the 10-core 10900K.

It depends upon what you want to do with it. If you are a programmer and want to experiment with AVX-512 you don't have much options. For running a multi-core chess engine the 10 core 10900K is probably better, is is cheaper and draws less power. Performance wise they are more or less on par. The 10900K has 2 extra cores and the 11900K has a higher IPC, not much of a difference.

If you don't need AVX-512 you are better of with the AMD 5950X, multi-core it has more than twice the speed of the 11900K, it is more expensive though.

[Modern Times]

Or indeed the 12-core 5900X