Commercial Release of Ethereal 13.00 (NNUE) for AVX2 Systems

[AndrewGrant]

No no, not at all. Its all of the above, not just the data. So the only to NOT be trained on other data while also NOT using the Stockfish pipeline while also NOT using the Stockfish implementation while also BEING high level.

I am trying to implement it in Booot just now. With Delphi+Python+Keras+Tensorflow. Not using any Stockfish pipeline. I am still not sure i can pass this way by my own, but today i am much closer to finish then i was 1 month ago before i started .

That's great to hear! The OpenBench discord is always a good place to bounce ideas off of people. Couple different authors with their own NNUE approaches.

[willmorton]

No no, not at all. Its all of the above, not just the data. So the only to NOT be trained on other data while also NOT using the Stockfish pipeline while also NOT using the Stockfish implementation while also BEING high level. I have no doubt that the other networks are trained on the data their authors claim. Sorry if there was any confusion there.

ok, i get your point

[AlexChess]

"Ethereal 13 (NNUE) includes two NNUE Networks for evaluation. These Networks are the second of their kind, boasting themselves as the only other high level NNUEs not derived from, trained on, nor duplicated from the works of the Stockfish team."
I'll stick to that. I'm the last person doing it on my own.
And food for thought: Stockfish folks training on Leela data is far less concerning to me than other engines reusing the entire SF pipeline.

We respect your decision. Is there any chance of us seeing an unofficial version with the feature to use large networks?

SCOOP:
Ethereal 13.00 NNUE github works with Fat Fritz 2.0 NET, so we can test it deeply



ALL THE 128 GAMES PLAYED UNTIL NOW:

https://onedrive.live.com/?cid=B0F79706 ... 69&o=OneUp

Blitz 3 mins + 3 secs /all between 8 top Engines on Mac mini M1 4 CPUs NATIVE 128/280 games played

I have tested Ethereal 13.00 standard and with 40 MB Fat Fritz 2 NNUE ; Ethereal 12.75 + Mike Byrne's nn-c4c306d60d95.nnue getting approximately the same performances. Most testers (and me too) have reported that Ethereal 13.00 NNUE commercial is 20 ELO stronger than Igel 3.0.5 NNUE, so with Stockfish and Fat Fritz 2.0 NNUEs the github versions seems stronger on my test...

My previous blitz between 38 top engines confirms my new results: http://talkchess.com/forum3/viewtopic.p ... 39#p896052

Notes:

-Ethereal (github) is very fast and calculates 10-15 mN/s. It is very strong, but still weaker than Stockfish & derivatives.
-It is much stronger than Igel 3.0.5 NNUE (I now replace it with Stockfish derivate ShashChess 17.1 Neon, restarting the games)
-It is a little weaker than Dragon 2.0

[pohl4711]

Most testers (and me too) have reported that Ethereal 13.00 NNUE commercial is 20 ELO stronger than Igel 3.0.5 NNUE,

I dont. In my testings Ethereal 13 nnue is +59 Elo stronger than Igel 3.0.5 nnue. Both engines have played 8000 games. So, the errorbar is +/-6 Elo, only, for both engines.

I think +75 Elo, in my ratinglist, of Ethereal 13 nnue compared to Ethereal 12.75 is a valid result, when mentioning, that A.Grants selfplay progression test was:

ELO | 121.51 +- 5.35 (95%)
CONF | 10.0+0.1s Threads=1 Hash=8MB
Games | N: 10436 W: 4966 L: 1458 D: 4012

So, in my testruns, using longer thinking-time (3'+1'' (but using slower hardware (Notebook, running engines on 20 Hyperthreading-Threads on a 12 core machine)), I would have expected an Elo-gain of around +70 to +80 Elo. And thats exactly, what we see.
The +45 or so Elo-gain of other testers seem pretty strange to me.

https://www.sp-cc.de

[AlexChess]

My complete tests are about Ethereal 13.0.0 github because commercial version doesn't work anymore on my Windows 10 ARM64 running under Parallels Desktop for mac M1 . It doesn't crash only if I use 1 CPU, (12.75 works perfectly on it sice 6 months) . 500 kN/s are unfair against the other engines (2-4 mN/s) . On mac OS M1 native it runs at 10-15 mN/s, but it is still the github version. Here the spread also of the github standard version against Igel 3.0.5 NNUE Neon is huge (258 ELO ) despite this last one scores well while playing against all others 30 M1 Neon engines!

Regards, AlexChess

[Madeleine Birchfield]

SCOOP:
Ethereal 13.00 NNUE github works with Fat Fritz 2.0 NET, so we can test it deeply



ALL THE 128 GAMES PLAYED UNTIL NOW:

https://onedrive.live.com/?cid=B0F79706 ... 69&o=OneUp

Blitz 3 mins + 3 secs /all between 8 top Engines on Mac mini M1 4 CPUs NATIVE 128/280 games played

I have tested Ethereal 13.00 standard and with 40 MB Fat Fritz 2 NNUE ; Ethereal 12.75 + Mike Byrne's nn-c4c306d60d95.nnue getting approximately the same performances. Most testers (and me too) have reported that Ethereal 13.00 NNUE commercial is 20 ELO stronger than Igel 3.0.5 NNUE, so with Stockfish and Fat Fritz 2.0 NNUEs the github versions seems stronger on my test...

My previous blitz between 38 top engines confirms my new results: http://talkchess.com/forum3/viewtopic.p ... 39#p896052

Notes:

-Ethereal (github) is very fast and calculates 10-15 mN/s. It is very strong, but still weaker than Stockfish & derivatives.
-It is much stronger than Igel 3.0.5 NNUE (I now replace it with Stockfish derivate ShashChess 17.1 Neon, restarting the games)
-It is a little weaker than Dragon 2.0

I do wonder how strong the github version of Ethereal 13.00 is with dkappe's Night Nurse net.

[CMCanavessi]

Hey Andrew, I got 2 loses on time in 8 games (so far), using Cute Chess, FRC, 3 min + 2 sec, 1 thread, ponder on, 1024 hash, non-pext avx2 binary. Tournament is still ongoing in my stream if you wanna take a look. Ethereal was comfortably winning both games against Shredder 13 and ended up losing on time. After that, it didn't happen again.

Also, a different thing: i was benchmarking avx vs avx2 vs ssse3 (all non-pext) with the go infinite command, all 3 at the same time and looking at the output. All 3 were showing the same ammount of nodes for every depth it was printing, until around depth 30 where avx2 started doing things different. avx and ssse3 kept printing the same node count, but avx2 was different. Is this behaviour normal?

[CMCanavessi]

Hey Andrew, I got 2 loses on time in 8 games (so far), using Cute Chess, FRC, 3 min + 2 sec, 1 thread, ponder on, 1024 hash, non-pext avx2 binary. Tournament is still ongoing in my stream if you wanna take a look. Ethereal was comfortably winning both games against Shredder 13 and ended up losing on time. After that, it didn't happen again.

Also, a different thing: i was benchmarking avx vs avx2 vs ssse3 (all non-pext) with the go infinite command, all 3 at the same time and looking at the output. All 3 were showing the same ammount of nodes for every depth it was printing, until around depth 30 where avx2 started doing things different. avx and ssse3 kept printing the same node count, but avx2 was different. Is this behaviour normal?

Another loss on time, 3rd in 10 games, this time Vs. SlowChessBlitz 2.5
I have now replaced the avx2 binary with the avx one. We'll see what happens and if that fixes it.

[Modern Times]

Hey Andrew, I got 2 loses on time in 8 games (so far), using Cute Chess, FRC, 3 min + 2 sec, 1 thread, ponder on, 1024 hash, non-pext avx2 binary. Tournament is still ongoing in my stream if you wanna take a look. Ethereal was comfortably winning both games against Shredder 13 and ended up losing on time. After that, it didn't happen again.

It is probably not an FRC problem specifically as I played over 3,000 FRC games without a single time loss. I wonder if it might be a ponder-on issue as most testers use ponder-off.

[CMCanavessi]

Hey Andrew, I got 2 loses on time in 8 games (so far), using Cute Chess, FRC, 3 min + 2 sec, 1 thread, ponder on, 1024 hash, non-pext avx2 binary. Tournament is still ongoing in my stream if you wanna take a look. Ethereal was comfortably winning both games against Shredder 13 and ended up losing on time. After that, it didn't happen again.

It is probably not an FRC problem specifically as I played over 3,000 FRC games without a single time loss. I wonder if it might be a ponder-on issue as most testers use ponder-off.

Might be, but I've been using Ethereal + ponder on for a long time without issues (on non-FRC games, though). Still there's something wrong, as ethereal is the only engine so far out of 20 that has lost on time... we'll see how it goes now that I've replaced avx2 with avx.