Introducing Igel chess engine

[AlexChess]

If you could compile Igel 3.0.0 Neon also for macOS M1 native, I'll be very happy. I alternate the tests on all 3 operating systems using Arena (Win) , BanksiaGUI (Mac) and Cutechess (Ubuntu ARM64)

It is an interesting challenge because I have no Apple devices. I will see what I can do.

These are the benchmarks of Igel 3.0 NNUE compiled by Acepoint on macOS BIg Sur Silicon M1 native, using 4 CPUs: It's 2 times faster than on Windows 10 ARM64 emulated on Parallels (always with 4 CPUs) 7000k starting position, 9000k changed queens (middle game) 12.000k on endgame.

[Krzysztof Grzelak]

Question to the author of the engine: what happens to this chess engine. That the engine does not start.

[voffka]

These are the benchmarks of Igel 3.0 NNUE compiled by Acepoint on macOS BIg Sur Silicon M1 native, using 4 CPUs: It's 2 times faster than on Windows 10 ARM64 emulated on Parallels (always with 4 CPUs) 7000k starting position, 9000k changed queens (middle game) 12.000k on endgame.

Nice one. I would only stick to native run and avoid using the virtualization

[voffka]

Hello Krzysztof,

Question to the author of the engine: what happens to this chess engine. That the engine does not start.

I would check the basics: if you are using the right version of Igel for your CPU: the official binaries support AVX2 and AVX512, so you need to check if your CPU supports that.

[Krzysztof Grzelak]

My CPU data is here: http://talkchess.com/forum3/viewtopic.php?f=6&t=77071

Please take a look at it:



I am using GUI Cutochess 1.2 by Guenther Simon.

[voffka]

My CPU data is here: http://talkchess.com/forum3/viewtopic.php?f=6&t=77071

Your CPU is from 2009, it does not have a support for AVX2, only SSE. I will try to compile tonight a build of Igel specifically targeted for your ancient CPU

[Krzysztof Grzelak]

Warmly he thanks and I am waiting on new versions.

[voffka]

Custom compiles of Igel 3.0.0 specifically targeted for their CPUs provided to Rebel and Krzysztof via PM.

[voffka]

Igel 3.0.5 at https://github.com/vshcherbyna/igel/releases/tag/3.0.5 (official executable binaries for Windows and IGN net).

This release brings significant improvements in search, time management and evaluation. I accidentally found a strong patch in search which brought around 12 elo in LTC in one shot. Overall strength increase due to search/time management improvements is around 20 elo, strength increase due to new network is 20 elo more. There are also some improvements in SMP mode that were quite good in short time controls: 15 elo, but hard to test on long time control as it will require a lot of computational power.

The new network ign-1-d593efbd is trained from scratch using nnue-pytorch on 12 billions of d8 positions using data from Igel 2.6.0 (HCE), Igel 2.9.0 (NNUE) and Igel 3.0.0 (NNUE) with a validation set of 10 million of d16 data generated using previous network ign-1-139b702b.

Enormous amount of computational power has been used to test this build on OpenBench, I would like to thank Andrew Grant and Bojun Guo (noobpwnftw) for supporting Igel project!

I would like to also thank authors of nnue-pytorch project for helping to make this release possible and more or less 'hassle free' training experience!

What's new:

- Train a new network using Igel 2.6.0 (HCE), Igel 2.9.0 (NNUE) and Igel 3.0.0 (NNUE) data: ign-1-d593efbd
- Better time management in sd time control
- Skip tt cutoff for null moves in search
- Skip null move prunning on tt hit
- Vote for best thread in SMP mode
- Less aggressive time use in non repeated tc

Regression run against Igel 3.0.0:

Long Time Control No Increment

Code: Select all

ELO   | 51.95 +- 2.85 (95%)
SPRT  | 60.0+0.0s Threads=1 Hash=64MB
LLR   | 0.00 (-2.94, 2.94) [0.00, 0.00]
Games | N: 20548 W: 5296 L: 2246 D: 13006

Long Time Control With Increment

Code: Select all

ELO   | 40.20 +- 2.74 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 0.00 (-2.94, 2.94) [0.00, 0.00]
Games | N: 20408 W: 4600 L: 2249 D: 13559

[Gabor Szots]

Thanks Volodymyr.

I understand you have to create a weights folder under src to compile it with an embedded net. But how do you compile it if you want to use an external net? In other words, what modifications shall I make to the makefile (I tried some to no avail)?