Commercial Release of Ethereal 13.00 (NNUE) for AVX2 Systems

[AndrewGrant]

Are you emulating something? That might be the worst NPS values I've ever seen. Ethereal does not work under WINE as there is no need if you are on Linux you would compile your own.

Yes, it is under Parallels Desktop for mac M1 and I was running a tournament on BanksiaGUI using all the power. Closed all now

Now that everything is closed, can you try your GUI again? Ethereal checks the time every 1024 nodes. If you were getting 60 nodes per second, and playing quick games, I can easily see a time loss with no move returned, due to the extremely low speed.

Ok, thank you for your patience. On BanksiaGui I was running a blitz with ethereal 12.75
+ Mike Byrne's NNUE-c4c3 and it is very strong... 13.00 after a while stopped thinking without crashing. I have always used Ethereal with it with no problems.

I don't follow. Did you attempt to run a new tournament in your GUI with Ethereal 13.00 now that you have stopped your existing workloads that were taking up your entire CPU? If so, did Ethereal function? If not, can you provide the UCI logs to me, and we can talk further via talkchess PMs or my email (andrew@grantnet.us)

[AlexChess]

Ethereal 13.00 was one of the 38 engines playing the tournament. Ethereal 12.75 is doing very well (crosstable on previous post)
with 4 CPU all was running fine and fast. When i started Ethereal 13.00 ssse3 games, with different opponents, it started fine, thinking for a while, then stopped calculating and loses all the games by time. If I run it on infinite level, with Banksiagui or Arena, it works fine.
The only engine that has the same issue is Pedone 2.1, 3.0, 3.1. I have contacted the author Fabio Gobbato and he kindly tried to fix it, but no joy.

I will search the engine logs and post them to the email that you have given to me. It doesn't lose by time during the game, it never makes the first move...

Thank you,
Regards, AlexChess

[Paloma]

Are you emulating something? That might be the worst NPS values I've ever seen. Ethereal does not work under WINE as there is no need if you are on Linux you would compile your own.

Yes, it is under Parallels Desktop for mac M1 and I was running a tournament on BanksiaGUI using all the power. Closed all now

Now that everything is closed, can you try your GUI again? Ethereal checks the time every 1024 nodes. If you were getting 60 nodes per second, and playing quick games, I can easily see a time loss with no move returned, due to the extremely low speed.

I have the same output in the command line as Alex.
i7 6700 3.6MHz, Ethereal-pext-avx.exe:

go wtime 10000
info depth 1 seldepth 1 multipv 1 score cp 5 time 0 nodes 3 nps 3000 tbhits 0 hashfull 0 pv a2a3
info depth 2 seldepth 2 multipv 1 score cp 72 time 0 nodes 19 nps 19000 tbhits 0 hashfull 0 pv g1f3 g8h6
info depth 3 seldepth 3 multipv 1 score cp 5 time 0 nodes 63 nps 63000 tbhits 0 hashfull 0 pv g1f3 g8f6 a2a3
info depth 4 seldepth 4 multipv 1 score cp 4 time 0 nodes 111 nps 111000 tbhits 0 hashfull 0 pv g1f3 g8f6 a2a3 b8c6
info depth 5 seldepth 6 multipv 1 score cp -1 time 0 nodes 371 nps 371000 tbhits 0 hashfull 0 pv c2c4 b8c6 a2a3 e7e6 g1f3
info depth 6 seldepth 4 multipv 1 score cp 30 time 0 nodes 729 nps 729000 tbhits 0 hashfull 0 pv c2c4 g8f6 b1c3 b8c6
info depth 7 seldepth 7 multipv 1 score cp 32 time 16 nodes 3343 nps 196000 tbhits 0 hashfull 0 pv c2c4 d7d5 c4d5 g8f6 b1c3 f6d5
info depth 8 seldepth 10 multipv 1 score cp 27 time 16 nodes 7493 nps 440000 tbhits 0 hashfull 2 pv c2c4 e7e6 e2e3 b8c6 g1f3 d7d5 c4d5 e6d5
info depth 9 seldepth 11 multipv 1 score cp 36 time 16 nodes 9955 nps 585000 tbhits 0 hashfull 3 pv c2c4 e7e6 g1f3 g8f6 b1c3 d7d5 c4d5 e6d5 e2e3
info depth 10 seldepth 14 multipv 1 score cp 37 time 16 nodes 15883 nps 934000 tbhits 0 hashfull 4 pv c2c4 g8f6 b1c3 d7d5 d2d4 d5c4 g1f3 c8g4 e2e3 c7c6 f1c4
info depth 11 seldepth 16 multipv 1 score cp 37 time 47 nodes 44084 nps 918000 tbhits 0 hashfull 9 pv d2d4 g8f6 c2c4 d7d5 b1c3 d5c4 g1f3 c8g4 e2e3 c7c6 f1c4
info depth 12 seldepth 17 multipv 1 score cp 30 time 63 nodes 86856 nps 1357000 tbhits 0 hashfull 15 pv d2d4 d7d5 g1f3 c8f5 e2e3 g8f6 c2c4 e7e6 b1c3 b8c6 f1e2 d5c4 e2c4
info depth 13 seldepth 23 multipv 1 score cp 29 time 109 nodes 164636 nps 1496000 tbhits 0 hashfull 23 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 b8c6 c1g5 f7f6 g5f4 c6b4 a1c1 c7c6 d1b3 d5c4 b3c4
info depth 14 seldepth 19 multipv 1 score cp 21 time 188 nodes 307406 nps 1626000 tbhits 0 hashfull 44 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 g8f6 c1f4 b8d7 c4d5 e6d5 d1b3
info depth 15 seldepth 24 multipv 1 score cp 19 time 219 nodes 395633 nps 1798000 tbhits 0 hashfull 58 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 g8f6 c1g5 c7c6 h2h3 h7h6 g5f6 d8f6 e2e3 a7a6 d1b3
info depth 16 seldepth 24 multipv 1 score cp 27 time 328 nodes 599193 nps 1821000 tbhits 0 hashfull 94 pv d2d4 d7d5 c2c4 d5c4 e2e3 c8e6 b1c3 g8f6 e3e4 c7c6 a2a3 e6g4 g1f3 g4f3 g2f3
bestmove d2d4 ponder d7d5

[AlexChess]

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

Yes agreed. The pure list never went away on FRC, but there is some sort of bug which means we had to turn them off for the other two lists. I'm trying to fix that for 40/15. One big issue with them is the number of games. Yes you reduce potential bias from testing against multiple versions of the same family, but you increase significantly in some cases the statistical margins of error because of less games.

For example Cfish is Stockfish optimized in C running 30% faster. ShashChess 17.1 implements Alexander Shashin mathematical theory and it's very strong. Corchess is perfect for long time analysis. But also the free (derivatives, NOT clones or they would score exactly like SF) Fisherov 0.98, Zeus 11, Eman 7.22, AI 15.00, KillFish CTR Hybrid v1.4 are deprecated because they do not release source being based on Stockfish GPL 3.0 (exactly like commercials Fat fritz 2 and Houdini 6 before they have been forced to drop the source code in github by the Stockfish team and fans) ...but often they score better than latest Stockfish 13-dev. Not considering them brings to a very partial list of all top engines. [Source] https://chessengines.blogspot.com/p/rating-jcer.html ...For me the most complete and trusted computer rating list *not 3600 Elo but 3000-3200 ELO* for top engines.

[AndrewGrant]

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

Yes agreed. The pure list never went away on FRC, but there is some sort of bug which means we had to turn them off for the other two lists. I'm trying to fix that for 40/15. One big issue with them is the number of games. Yes you reduce potential bias from testing against multiple versions of the same family, but you increase significantly in some cases the statistical margins of error because of less games.

For example Cfish is Stockfish optimized in C running 30% faster. ShashChess 17.1 implements Alexander Shashin mathematical theory and it's very strong. Corchess is perfect for long time analysis. But also the free (derivatives, NOT clones or they would score exactly like SF) Fisherov 0.98, Zeus 11, Eman 7.22, AI 15.00, KillFish CTR Hybrid v1.4 are deprecated because they do not release source being based on Stockfish GPL 3.0 (exactly like commercials Fat fritz 2 and Houdini 6 before they have been forced to drop the source code in github by the Stockfish team and fans) ...but often they score better than latest Stockfish 13-dev. Not considering them brings to a very partial list of all top engines. [Source] https://chessengines.blogspot.com/p/rating-jcer.html ...For me the most complete and trusted computer rating list *not 3600 Elo but 3000-3200 ELO* for top engines.

You seriously over estimate the amount of effort that goes into those Stockfish forks. The only "fork" worth anyone's time is CFish, because its simply faster. All of the other ones are gimicks, elo losing changes, merging PRs faster than SF team, or merging yellow LTCs. People who are capable of increasing the strength of Stockfish are contributing to Stockfish. They are not spinning their own forks.

[AlexChess]

Are you emulating something? That might be the worst NPS values I've ever seen. Ethereal does not work under WINE as there is no need if you are on Linux you would compile your own.

Yes, it is under Parallels Desktop for mac M1 and I was running a tournament on BanksiaGUI using all the power. Closed all now

Now that everything is closed, can you try your GUI again? Ethereal checks the time every 1024 nodes. If you were getting 60 nodes per second, and playing quick games, I can easily see a time loss with no move returned, due to the extremely low speed.

I have the same output in the command line as Alex.
i7 6700 3.6MHz, Ethereal-pext-avx.exe:

go wtime 10000
info depth 1 seldepth 1 multipv 1 score cp 5 time 0 nodes 3 nps 3000 tbhits 0 hashfull 0 pv a2a3
info depth 2 seldepth 2 multipv 1 score cp 72 time 0 nodes 19 nps 19000 tbhits 0 hashfull 0 pv g1f3 g8h6
info depth 3 seldepth 3 multipv 1 score cp 5 time 0 nodes 63 nps 63000 tbhits 0 hashfull 0 pv g1f3 g8f6 a2a3
info depth 4 seldepth 4 multipv 1 score cp 4 time 0 nodes 111 nps 111000 tbhits 0 hashfull 0 pv g1f3 g8f6 a2a3 b8c6
info depth 5 seldepth 6 multipv 1 score cp -1 time 0 nodes 371 nps 371000 tbhits 0 hashfull 0 pv c2c4 b8c6 a2a3 e7e6 g1f3
info depth 6 seldepth 4 multipv 1 score cp 30 time 0 nodes 729 nps 729000 tbhits 0 hashfull 0 pv c2c4 g8f6 b1c3 b8c6
info depth 7 seldepth 7 multipv 1 score cp 32 time 16 nodes 3343 nps 196000 tbhits 0 hashfull 0 pv c2c4 d7d5 c4d5 g8f6 b1c3 f6d5
info depth 8 seldepth 10 multipv 1 score cp 27 time 16 nodes 7493 nps 440000 tbhits 0 hashfull 2 pv c2c4 e7e6 e2e3 b8c6 g1f3 d7d5 c4d5 e6d5
info depth 9 seldepth 11 multipv 1 score cp 36 time 16 nodes 9955 nps 585000 tbhits 0 hashfull 3 pv c2c4 e7e6 g1f3 g8f6 b1c3 d7d5 c4d5 e6d5 e2e3
info depth 10 seldepth 14 multipv 1 score cp 37 time 16 nodes 15883 nps 934000 tbhits 0 hashfull 4 pv c2c4 g8f6 b1c3 d7d5 d2d4 d5c4 g1f3 c8g4 e2e3 c7c6 f1c4
info depth 11 seldepth 16 multipv 1 score cp 37 time 47 nodes 44084 nps 918000 tbhits 0 hashfull 9 pv d2d4 g8f6 c2c4 d7d5 b1c3 d5c4 g1f3 c8g4 e2e3 c7c6 f1c4
info depth 12 seldepth 17 multipv 1 score cp 30 time 63 nodes 86856 nps 1357000 tbhits 0 hashfull 15 pv d2d4 d7d5 g1f3 c8f5 e2e3 g8f6 c2c4 e7e6 b1c3 b8c6 f1e2 d5c4 e2c4
info depth 13 seldepth 23 multipv 1 score cp 29 time 109 nodes 164636 nps 1496000 tbhits 0 hashfull 23 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 b8c6 c1g5 f7f6 g5f4 c6b4 a1c1 c7c6 d1b3 d5c4 b3c4
info depth 14 seldepth 19 multipv 1 score cp 21 time 188 nodes 307406 nps 1626000 tbhits 0 hashfull 44 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 g8f6 c1f4 b8d7 c4d5 e6d5 d1b3
info depth 15 seldepth 24 multipv 1 score cp 19 time 219 nodes 395633 nps 1798000 tbhits 0 hashfull 58 pv d2d4 d7d5 g1f3 c8f5 c2c4 e7e6 b1c3 g8f6 c1g5 c7c6 h2h3 h7h6 g5f6 d8f6 e2e3 a7a6 d1b3
info depth 16 seldepth 24 multipv 1 score cp 27 time 328 nodes 599193 nps 1821000 tbhits 0 hashfull 94 pv d2d4 d7d5 c2c4 d5c4 e2e3 c8e6 b1c3 g8f6 e3e4 c7c6 a2a3 e6g4 g1f3 g4f3 g2f3
bestmove d2d4 ponder d7d5

Considering that I'm running Ethereal 13.00 ssse3 using only 4CPUs under Parallels desktop 16.5 M1 Windows 10 ARM64 my nodes 395633 nps 1098000 are very good! Thank you for sharing. Now I'm running it on my brother's Ryzen 3900x 24 CPU, because I have issues on my hardware. (but Ethereal 12.75 NNUE works perfectly on it since 7 months...)

Regards, AlexChess

[AlexChess]

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

Yes agreed. The pure list never went away on FRC, but there is some sort of bug which means we had to turn them off for the other two lists. I'm trying to fix that for 40/15. One big issue with them is the number of games. Yes you reduce potential bias from testing against multiple versions of the same family, but you increase significantly in some cases the statistical margins of error because of less games.

For example Cfish is Stockfish optimized in C running 30% faster. ShashChess 17.1 implements Alexander Shashin mathematical theory and it's very strong. Corchess is perfect for long time analysis. But also the free (derivatives, NOT clones or they would score exactly like SF) Fisherov 0.98, Zeus 11, Eman 7.22, AI 15.00, KillFish CTR Hybrid v1.4 are deprecated because they do not release source being based on Stockfish GPL 3.0 (exactly like commercials Fat fritz 2 and Houdini 6 before they have been forced to drop the source code in github by the Stockfish team and fans) ...but often they score better than latest Stockfish 13-dev. Not considering them brings to a very partial list of all top engines. [Source] https://chessengines.blogspot.com/p/rating-jcer.html ...For me the most complete and trusted computer rating list *not 3600 Elo but 3000-3200 ELO* for top engines.

You seriously over estimate the amount of effort that goes into those Stockfish forks. The only "fork" worth anyone's time is CFish, because its simply faster. All of the other ones are gimicks, elo losing changes, merging PRs faster than SF team, or merging yellow LTCs. People who are capable of increasing the strength of Stockfish are contributing to Stockfish. They are not spinning their own forks.

Most of them are also in the Stockfish Team.

OK, different opinions, but I still consider you a myth!

PS: Andrew, on Ryzen 3900x Ethereal 13.00 ssse3 works perfectly @40 mN/s , so the crashing issue is related only to my emulation, THANK YOU FOR THE SUPPORT GIVEN! ... I still have the wonderful Ethereal 12.75 NNUE to test.

[AndrewGrant]

PS: Andrew, on Ryzen 3900x Ethereal 13.00 ssse3 works perfectly @40 mNps , so the crashing issue is related only to my emulation, THANK YOU FOR THE SUPPORT GIVEN! ... I still have the wonderful Ethereal 12.75 NNUE to test.

If by Ethereal 12.75 NNUE (Etherlito, a play on years of people stealing engines) , you are referring to the version using a Stockfish Network, it will beat the pants off of 13.00. It will even beat SF11. A testament to the low effort required to reuse a Stockfish Network. But its not original, its not special, and no one should care for it. Likewise for other SF Network users.

If you just have the standard Ethereal 13.00, the AVX2 should beat your SSSE3 by a little bit. Even the "Standard" version actually has what you could call an "efficiently hashed non-updatable Neural Network". While it does not contain explicit vectorization, the compiler is doing it.

[AlexChess]

PS: Andrew, on Ryzen 3900x Ethereal 13.00 ssse3 works perfectly @40 mNps , so the crashing issue is related only to my emulation, THANK YOU FOR THE SUPPORT GIVEN! ... I still have the wonderful Ethereal 12.75 NNUE to test.

If by Ethereal 12.75 NNUE (Etherlito, a play on years of people stealing engines) , you are referring to the version using a Stockfish Network, it will beat the pants off of 13.00. It will even beat SF11. A testament to the low effort required to reuse a Stockfish Network.

If you just have the standard Ethereal 13.00, the AVX2 should beat your SSSE3 by a little bit. Even the "Standard" version actually has what you could call an "efficiently hashed non-updatable Neural Network". While it does not contain explicit vectorization, the compiler is doing it.

Yes, I'm surprised by the strengh using MikeB's nn-c4c3... from Stockfish NNUE Test repository. With banksiaGui I can choose games to play first, and it has scored 22/37 games!!!! (CFish 23.5/37) testing also 13.00 ssse3 with the same NNUE, if my brother doesn't return asking his PC. ...Happy for your return!

[Leto]

Hi Andrew,
It would be nice if you can put more buying options like paypal on your list...
When only Credit Card, it would seem you are trying to avoid paying taxes....

I do not have a credit card (and dont want to get one). So, without the possibility to pay with PayPal or GooglePay, I cannot buy Ethereal 13 and cannot test it.

I am in the process of getting Paypal to work. Originally I was planning to use Paypal, as it has greater name recognition, and charges lower fees on my end, where as Stripe takes a fair chunk for non US transactions. Paypal integration seemed like it was only meant for sending physical products, not making digital purchases. As a last resort, I might end up adding a Paypal link which will not be a "sale" page, but simply a money transfer page. Downside is that I will not be able to automate the process as well.

Paypal is used for many online transactions. I've been using paypal for years to pay my entrance fees for ICCF tournaments.