Ho ho ho it's tournament time.

[gcp]

To make it easier to make the tournament deadline,
get a submission-ready branch set up.

Merges most of the outstanding strength patches:

• Bad bishop eval (Bad bishop eval. #56).
• Late move pruning changes (Late move pruning based on quiets #78).
• Improving variable (Introduce improving variable #83).
• Include promotions in qsearch (Include promotions in Qsearch #97).

Also include #95 to make this exactly 4096 bytes,
and put in tournament settings (hash/threads) that keep the size at exactly that.

vs master (432c06a):

ELO | 18.58 +- 9.34 (95%)
SPRT | 8.0+0.08s Threads=1 Hash=64MB
LLR | 2.97 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 3032 W: 945 L: 783 D: 1304

[tissatussa]

when exactly does the tournament start and which link to visit ?

[gcp]

https://tcec-chess.com/
https://wiki.chessdom.org/TCEC_4k_Rules

The submission deadline is somewhere in the middle of Christmas day (estimated: 18:30:49 2022-12-25) when the current TCEC tournament finishes.

I understand @kz04px is working on an improved minifier that will allow some more outstanding strength patches to go in, but I wanted this set up, tested and ready in case something goes wrong.

[gcp]

The TCEC tournament machine is currently

ALL IN ONE BOX SYSTEM

    CPU: 2x Xeon 6230R (52 cores/104 threads)
    GPU: 2xA100-PCIE-40GB
    RAM: 256GiB (~96GiB/engine)
    Storage: 2TiB SSD + Starting from S23 VVLTC: 12TiB (4 drives) SSD for Syzygy3-7 (total 14TB) + 15TiB (2 drives raid1 (30TiB)) HDD for Syzygy7 DTZ
    OS: CentOS Stream release 8
    Drivers: nvidia-driver.x86_64 3:495.29.05-1.el8 cuda.x86_64 11.5.1-1 libcudnn8.x86_64 8.3.2.44-1.cuda11.5
    Compilers: gcc 8.5.0, 9.2.1, 10.3.1 and 11.2.1; clang 13.0.0; rust 1.58.1

It's possible we get weaker with more threads at a certain point, but I don't have any way of testing this. If it's only 52 cores I'll try to get our threads closer to that, will update PR if I succeed.

[gcp]

Some testing of 32 thread vs 16 thread on a 16c/32t machine ended up at -50 +- 25 in favor of using only the real cores, so I recommend we do not use more than 52 threads on the TCEC machine.

Scaling from 8c to 16c in itself (i.e. only using real cores) is fine:

Score of 16th vs 8th: 98 - 37 - 115  [0.622] 250
...      16th playing White: 58 - 15 - 52  [0.672] 125
...      16th playing Black: 40 - 22 - 63  [0.572] 125
...      White vs Black: 80 - 55 - 115  [0.550] 250
Elo difference: 86.5 +/- 31.8, LOS: 100.0 %, DrawRatio: 46.0 %

[tissatussa]

good luck with that tournament! I will follow it. I think version #100 has about rating 2600+, using 10m+3s for test. Which other articipants are there ?

[kz04px]

good luck with that tournament! I will follow it. I think version #100 has about rating 2600+, using 10m+3s for test. Which other articipants are there ?

https://github.com/ONE-RANDOM-HUMAN/we4k
https://github.com/scs-ben/pygone
https://github.com/MinusKelvin/ice4
There might be a fourth engine that is currently unknown.

[kz04px]

I've sent this version of the engine to the TCEC. Whether a newer submission is created in time remains to be seen.

[gcp]

Updated to https://github.com/kz04px/4ku/releases/tag/v1.0 for the tournament.