SF NNUE absolutely smokes on AMD

[Dann Corbit]

I just can't believe how well it runs. I put the two most recent Sergio nets for NNUE in the archive. I built it from the bleeding edge source.
It runs at SF speed on my AMD threadripper. At first, I thought I must have forgotten to connect the net, but no. The net is working,

Slack-jawed, buggy-eyed wowsers.

[zullil]

I just can't believe how well it runs. I put the two most recent Sergio nets for NNUE in the archive. I built it from the bleeding edge source.
It runs at SF speed on my AMD threadripper. At first, I thought I must have forgotten to connect the net, but no. The net is working,

Slack-jawed, buggy-eyed wowsers.

Seems hard to believe. I just built from the nnue-player-wip branch at the Stockfish repo. I ran bench with Use NNUE = false (the default) and then again with Use NNUE = true (and a recent sergio nn). The second bench took about twice as long as the first.

Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz

Linux 4.15.0-112-generic (Ubuntu 18.04.4 LTS)

What special instructions would make your AMD binary so much faster? Mine has bmi2 avx2

[towforce]

I've heard someone who bought a ThreadRipper CPU complain that people who write software should use more threading, because now that processor is in his PC, it doesn't have enough processing to do!

[Dann Corbit]

I just can't believe how well it runs. I put the two most recent Sergio nets for NNUE in the archive. I built it from the bleeding edge source.
It runs at SF speed on my AMD threadripper. At first, I thought I must have forgotten to connect the net, but no. The net is working,

Slack-jawed, buggy-eyed wowsers.

Seems hard to believe. I just built from the nnue-player-wip branch at the Stockfish repo. I ran bench with Use NNUE = false (the default) and then again with Use NNUE = true (and a recent sergio nn). The second bench took about twice as long as the first.

Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz

Linux 4.15.0-112-generic (Ubuntu 18.04.4 LTS)

What special instructions would make your AMD binary so much faster? Mine has bmi2 avx2

OK, the mystery is solved.
I was running as ordinary SF, and I based the "net is being used" on correct eval even from short searches.
I will run again with the checkbox checked and see how it goes.

[Dann Corbit]

OK, now that I am actually running the nnue version, I am getting 60% of Stockfish speed.


Still pretty good, but it was too good to be true.

[zullil]

OK, now that I am actually running the nnue version, I am getting 60% of Stockfish speed.


Still pretty good, but it was too good to be true.

[Zenmastur]

OK, now that I am actually running the nnue version, I am getting 60% of Stockfish speed.


Still pretty good, but it was too good to be true.

I did basically the same thing the other day. I was wondering which was the fastest EXE version for threadripper, so I started testing them all against the new PO versions. The version that was the fastest was the no.nnue version. Unfortunately it doesn't use a net, just straight SF. LOL

[Ovyron]

OK, now that I am actually running the nnue version, I am getting 60% of Stockfish speed.


Still pretty good, but it was too good to be true.

I did basically the same thing the other day. I was wondering which was the fastest EXE version for threadripper, so I started testing them all against the new PO versions. The version that was the fastest was the no.nnue version. Unfortunately it doesn't use a net, just straight SF. LOL

What I did when compiling NNUE for the first time was forgetting to set CPUs to 4, so I thought my compile was running at 1/10 of the speed that it should, and spent 2 days looking for a solution. Which was... to set threads=4 instead of default threads=1!

[towforce]

What I did when compiling NNUE for the first time was forgetting to set CPUs to 4, so I thought my compile was running at 1/10 of the speed that it should, and spent 2 days looking for a solution. Which was... to set threads=4 instead of default threads=1!

I would guess that if, like Dann, you have a CPU like ThreadRipper, the optimal setting for threads would be higher than 4.

[Dann Corbit]

I am running 60 threads.
That way, I can run three threads and 2 GPUs for LC0 at the same time, and have one thread left over for whatever.