AMD 9950x with Stockfish

[Vinvin]

New records for the new 16 cores CPU.
But not a lot of information about the new instruction set speed up : AVX512-VNNI, AVX256-VNNI, AVX512.



from https://www.phoronix.com/review/amd-ryz ... x-9900x/14

and


from https://www.techpowerup.com/review/amd- ... 50x/6.html

[smatovic]

Interesting.

Ipman benchmarks show SF 14.1 speedup of Zen 3 to Zen 4 as ~x1.3, now these benchmarks show ~x2.

https://ipmanchess.yolasite.com/amd--in ... ckfish.php

AFAIK SF 16.1 has bigger net size, and a second, smaller lazy network, so here maybe Amdahl's law steps in, cos newer SFs rely more on AVX compute?

Accoding to Wikipeda Zen 3 has 256b AVX2 (Idk how many pipelines), and Zen 4 AVX-512 ISA via 2x256b (four 256b pipelines), now Zen 5 with additional 512b data path.

Would make sense to me, that SF 16.1 can boost more via AVX-512 than 14.1, or alike.

--
Srdja

[h1a8]

What about the new intel Ultra 9 285k processor?
Any stockfish benchmarks for those?

[Werewolf]

Small improvement

[Jouni]

If You use Stockfish dev and run "speedtest" what is the result? My quess 15 to 20 Mnps or less. In TCEC Stockfish with 101 threads has 30 Mnps speed!

[Damir]

SF speed is reduced by more than 50% after NUMA implementation. NUMA implementation in SF is buggy and reduces SF speed by more than 50%.. I suppose SF team are unable to fix it? This has been an issue for months...

[cpeters]

Ipman benchmarks show SF 14.1

One year ago I didn't mind, but really, it's a three year old software. It measures just something now..

[Ras]

One year ago I didn't mind, but really, it's a three year old software. It measures just something now..

The point is not benchmarking Stockfish, but benchmarking CPUs. If you change the SF version, you cannot compare older benchmarks to newer ones anymore, any nobody is going to re-benchmark the older CPUs.

[cpeters]

Yes, exactly.

If you change the SF version, you cannot compare older benchmarks to newer ones anymore, any nobody is going to re-benchmark the older CPUs.

I had Srdja's statement in mind:

Ipman benchmarks show SF 14.1 speedup of Zen 3 to Zen 4 as ~x1.3, now these benchmarks [16.1] show ~x2.

Which I found more interesting as opposed to how fast old/new hardware runs old programs.

[Hai]

One year ago I didn't mind, but really, it's a three year old software. It measures just something now..

The point is not benchmarking Stockfish, but benchmarking CPUs. If you change the SF version, you cannot compare older benchmarks to newer ones anymore, any nobody is going to re-benchmark the older CPUs.

Stockfish developers have done a lot of changes and improvements and... to improve Stockfish on specific cpus since Stockfish 14.1.

The problem is:

Stockfish 14.1
AMD CPU 50000 kn/s
Intel CPU 50000 kn/s
Apple CPU 28000 kn/s

Newest Stockfish dev on the same CPU
AMD CPU 55000 kn/s
Intel CPU 51000 kn/s
Apple CPU 56000 kn/s

We don't need ipman to delete his list. It's a good list.
We need him to make a second list.