Commercial Release of Ethereal 13.00 (NNUE) for AVX2 Systems

[Wilson]

congratulations Andrew!
I am glad you went from "I don't know the first thing about nn and I'm willing to pay anybody to train a network for Ethereal" to doing your own thing! that's better than all the "whining for relevance" you did last year for sure.

Ethereal tends to lose to Stockfish and its derivatives (NNUE wise, or Fire / Houdini) more so than it loses to the rest of the AB field.

as for Fire I guess you mean its search as Mike B., Connor M. and Thomas M. all seem to agree on Fire's evaluation not being derived from Stockfish and an other well known forum member said that Fire's "evaluation is completely different from Stockfish" and that an other engine's evaluation he won't mention (the one with the long diagonal bishop code) is way more "fishy" than Fire's.

I wonder which engine he was referring to?!?

Fire's evaluation relationship with Houdini's is an other thing but you didn't mean that.

[Modern Times]

Only problem, I can imagine, is, to use the wrong (FRC) nnue-net for classical chess instead of the standard net.

Before my test-machine crashed (it is now for repair at XMG), more than 1000 of the 7000 games testrun of Ethereal 13 avx2 were played and I got a performance around +75 Elo compared to Ethereal 12.75.

Definitely the correct NET used so there is no problem.

Simply a difference in testing conditions, quite common and not just Ethereal. Larger error margins too with the smaller number of games due to time control.

[Modern Times]

After 700+games at 40/15 standard chess 1 thread, I have here +43 Elo per BayesElo, and +53 Elo according to Ordo. The latter being about what Andrew expected on our lists. The blitz games are still in progress - the improvement there could be a bit more.

[AndrewGrant]

After 700+games at 40/15 standard chess 1 thread, I have here +43 Elo per BayesElo, and +53 Elo according to Ordo. The latter being about what Andrew expected on our lists. The blitz games are still in progress - the improvement there could be a bit more.

Looks like we are in good shape then across the board.

Your FRC results so far place Ethereal at #3, with SF > K > E > H > F, as desired.
Your 40/15 results so far will comfortably put Ethereal above Houdini, and SF Network/implementation engines.
CEGT's early testing (200 games) has Ethereal at +83, so even if that drops a good bit its still in range.
And Pohl stated he is seeing +75 after ~1000 games, which would place Ethereal above (maybe) Dragon (MCTS), Komodo (14, non NNUE), and Houdini.

Very good results here.

[Modern Times]

Yes, here is Ethereal's position on the single CPU list:

http://ccrl.chessdom.com/ccrl/4040/cgi/ ... librate=no

and the individual performances:

http://ccrl.chessdom.com/ccrl/4040/cgi/ ... _00_64-bit

[AlexChess]

They aren't listed on CCRL because they're not unique engines. Rather, they're just a bunch of Stockfish derivatives which differ only slightly and will almost all play effectively identical to Stockfish with nearly identical evaluations.

I think that open source derivatives are very useful. They aren't simply clones and absolutely do not play the same moves of Stockfish. They contribute to greatly improve It and all the other Stockfish variations with new ideas. I'm very grateful to skilled programmers like Mike Byrne for the 5 strong variations and Andrea Manzo for ShashChess. They deserves my consideration and I'll continue to test every new build coming from them.

Regards, AlexChess

[AndrewGrant]

Yes, here is Ethereal's position on the single CPU list:

http://ccrl.chessdom.com/ccrl/4040/cgi/ ... librate=no

and the individual performances:

http://ccrl.chessdom.com/ccrl/4040/cgi/ ... _00_64-bit

Thanks. Its nice to see that Pure lists are back ( At least for FRC ) . That is what I always looked at back in the day. I'de love to see those for each list, and actually have that be the default. Its simply a greater quality list by removing sampling biases

I've finished adding support for AVX, and SSSE3 now. That means virtually all machines can now use Ethereal NNUE, although like with any engine, the SSSE3 versions will not make use of the 256-bit vector operations, and will see slowdowns.

Setting up PayPal has proved impossible. I tried two different vendors that are built on top of Paypal, and neither could do what was needed. PayPal itself was faltering, and was eating payments without producing any record or trace or charges. I'm going to have to stick with Stripe for now, which removes PayPal as an option. Perhaps I can deal with people on an individual basis over PayPal. PayPal as a merchant has not worked, but I recall PayPal being very good at simply sending money from point A to point B.

Once final results come in, I will make just one more thread to express that (virtually) all CPUs are not covered. Both in my own interest to get more eyes on Ethereal, and also since many have asked and that means many did not ask.

Also, the Standard binaries for the NNUE version (Windows) have been added to the github release. Which I will mention in whatever new thread in the coming days/weeks. Ipman did some early testing, and they show no more than say +7 elo over 12.75. This is two-fold: Firsly, NNUE was the major focus. Secondly, some components of Ethereal had to be tweaked to allow NNUE (Like not storing evals without regard for White/Black) which is a "slowdown" for non NNUE.

[mehmet123]

After 700+games at 40/15 standard chess 1 thread, I have here +43 Elo per BayesElo, and +53 Elo according to Ordo. The latter being about what Andrew expected on our lists. The blitz games are still in progress - the improvement there could be a bit more.

Which version of Ethereal 13 (Ethereal 13 or Ethereal 13 NNUE) do you use in your Standart Chess and FRC tests ?

[AndrewGrant]

After 700+games at 40/15 standard chess 1 thread, I have here +43 Elo per BayesElo, and +53 Elo according to Ordo. The latter being about what Andrew expected on our lists. The blitz games are still in progress - the improvement there could be a bit more.

Which version of Ethereal 13 (Ethereal 13 or Ethereal 13 NNUE) do you use in your Standart Chess and FRC tests ?

He is using Ethereal 13.00 NNUE

[Modern Times]

I'll make reference to NNUE in the engine notes. Also it needs to be coloured blue now since it is commercial.