Network Test (2) Dimension 2560 vs 2048

[Eduard Nemeth]

Network Test (2) Dimension 2560 vs 2048:

I did another test today. Already in my first test, the new unofficial dimension 2560 won against the SF dev standard network with dimension 2048. My private engine Smile XPro is not identical to SF dev. It is a modified SF version, and adapted to the new NNUE Dimension 2560.

SF dev used network "nn-c38c3d8d3920.nnue"
Smile XPro used network "nn-1ceb1ade2048.nnue"

After 36 games, Smile XPro was able to achieve a +4 lead. So it is clear to me that the new NNUE net with dimension 2560 plays better than SF dev when played with enough time. Bullet games often end randomly, in 10 minute games the engines take more time (sometimes even more than 60s per move). Ponder ON is very important to me, because we play with it on all servers.

The games was played on Ryzen 2700 with the Fritz 18 GUI.

6 Threads/Engine
Ponder ON
Timecontrol 10m + 0.1s
Hash 512 MB/Engine
Book EN-Tournament with color swap
Syzygy all 3456men

Result:


All games are here as PGN, see for yourself:
https://pixeldrain.com/u/nZeChHBk

Live picture:


My Homepage:
https://solistachess.jimdosite.com/

[Anton101]

Many thanks for test and the games!

Besty regards.

A. Ces

[Eduard Nemeth]

If you optimize the engine's selectivity to the larger network, then the engine will play better than with the smaller network. Probably not yet in Bullet, but if a reasonable search depth is reached, then the large network evaluate better. Unfortunately, the network with dimension 3072 plays weaker again (for me). I will no longer be using Dimension 2048 for my engines, this is the second time that Dimension 2560 has won a test. I'm not going back to smaller networks anymore. Also in position tests the larger network is better (if the engine is optimized for the network), and live on PlayChess.com it also plays very well, even on slow hardware.

[Anton101]

I have Cool Iris which is CorChess but with the experience file, so I am going to include that network, I will follow your advice.

Thank you so much.

Best regards.

[Homayoun]

Network Test (2) Dimension 2560 vs 2048:

I did another test today. Already in my first test, the new unofficial dimension 2560 won against the SF dev standard network with dimension 2048. My private engine Smile XPro is not identical to SF dev. It is a modified SF version, and adapted to the new NNUE Dimension 2560.

SF dev used network "nn-c38c3d8d3920.nnue"
Smile XPro used network "nn-1ceb1ade2048.nnue"

After 36 games, Smile XPro was able to achieve a +4 lead. So it is clear to me that the new NNUE net with dimension 2560 plays better than SF dev when played with enough time. Bullet games often end randomly, in 10 minute games the engines take more time (sometimes even more than 60s per move). Ponder ON is very important to me, because we play with it on all servers.

The games was played on Ryzen 2700 with the Fritz 18 GUI.

6 Threads/Engine
Ponder ON
Timecontrol 10m + 0.1s
Hash 512 MB/Engine
Book EN-Tournament with color swap
Syzygy all 3456men

Result:


All games are here as PGN, see for yourself:
https://pixeldrain.com/u/nZeChHBk

Live picture:


My Homepage:
https://solistachess.jimdosite.com/

Thank you very much for the test. It seems that in future stockfish will be in the same direction as leela. The nnue net will be larger and larger in size. Larger net means more knowledge. The difference , leela will be a Gpu engine and stockfish will be a CPU engine.

[Eduard Nemeth]

The size of the network with dimension 2560 is 61 MB. For a few weeks, however, the networks have been compressed. If the network were not compressed, it would be around 120 MB in size. The first NNUE networks had a volume of around 20 MB and were not compressed. I guess the NNUE networks will continue to grow, there is already a test network with dimension 3072.