TCEC S29 division P started

[smatovic]

The AMD architecture allows transparent access to the video RAM and system RAM to both the CPUs and GPUs. They have already implemented it for their AI workhorse type GPUs. I think Nvidia is doing something similar. Maybe Srdja can comment.

The big problem with commodity GPUs is that they spend all their time copying work to and from video RAM

Leela isn’t memory bandwidth bottlenecked. It does matter but not as much as you think it does. Leela sends work to each gpu, so nvlink really doesn’t do much.

Yes, true, Leela runs batches of parallel threads to utilize the GPU, and the neural network for evaluation is stored in GPU's VRAM. But I guess Dann was up to some kind of heterogeneous engine, utilizing CPU/GPU/NPU all together on the same game tree stored in unified memory. Idk how such an approach could look like, but Apple M-series offers already the needed hardware.

Lc0's batch based parallel approach is needed because of the host-device-latencies for kernel launch, not because of memory copies.

Host-Device Latencies
https://www.chessprogramming.org/GPU#Ho ... _Latencies

Idk what the kernel-launch overhead looks like on Apple M-series hardware.

--
Srdja

[Norm Pollock]

Final is running now.

The S29 SuperFinal is turning out to be a disaster for Reckless. SF has won 7 out of 7 as White. But 2 of those games also had Reckless winning in the Reverse Match as White, in effect making each of those 2 games 1/2-1/2. Adjusting for that, SF has won 5 out of 5 as White and in the 5 Reverse matches, Reckless has drawn as White.

Both Reckless and SF have been updated for the SuperFinal. Either SF had a great update, or Reckless' update backfired, or both. Engines cannot be modified mid-tournament except perhaps for crashes, but it would be nice if Reckless could substitute its Division P version, which was effective vs the previous Stockfish version.

[Uri Blass]

Final is running now.

The S29 SuperFinal is turning out to be a disaster for Reckless. SF has won 7 out of 7 as White. But 2 of those games also had Reckless winning in the Reverse Match as White, in effect making each of those 2 games 1/2-1/2. Adjusting for that, SF has won 5 out of 5 as White and in the 5 Reverse matches, Reckless has drawn as White.

Both Reckless and SF have been updated for the SuperFinal. Either SF had a great update, or Reckless' update backfired, or both. Engines cannot be modified mid-tournament except perhaps for crashes, but it would be nice if Reckless could substitute its Division P version, which was effective vs the previous Stockfish version.

Ido not know if it is the reason for the bad result in the first games but there was a problem with the speed of reckless and the games are going to be restarted.

Note also that even in case there was no problem in reckless your conclusion that either SF had a great update or reckless's update backfired is wrong and it is clearly possible that in the first 14 games stockfish is leading 9.5-4.5 when the final result is only 50-50(not that I expect it and I expect stockfish to get better result but I remember Junior was leading 5-0 against Fritz in a match about the right to play kramnik when the final result was 12-12 and later Fritz won to play against Kramnik.

5 games are nothing and can be a statistical mistake or simply positions when one engine is better.

[Norm Pollock]

I cannot find what the exact malfunction was that caused TCEC to restart the superfinal. I even checked with ChatGPT. I am glad that TCEC did restart it, because something was clearly wrong.

Regardless of whether this has anything to do with the malfunction, I think that the version numbering system for Reckless is confusing.

Based on their GitHub page: https://github.com/codedeliveryservice/Reckless

their prior version numbers are: v0.1.0,v0.2.0. v0.3.0, v0.4.0, v0.5.0, v0.6.0, v0.7.0, v0.8.0, v0.9.0, and their latest version is v0.10.0.

In computer science the "." is a delimiter, but in arithmetic it is a decimal point. Could anyone get confused between v0.1.0 and v0.10.0 ?

Just suppose the operator loaded v0.1.0 instead of v0.10.0, that would explain both the decrease in nps and the poor results.

Please do not take this as a fact, it is just a supposition of what might have caused the malfunction, which, so far, has not been made public.

[Joerg Oster]

https://www.chessdom.com/when-code-meet ... fair-play/

[Norm Pollock]

Reckless compiled with a new version of the Rust compiler and then submitted the binary to TCEC? Was the new version tested? I would think that a significant decrease in nps and inferior results would have been noticed in any reasonably sized test. Smells fishy (pun intended).

[Jouni]

This faster compile really makes difference! After 12 games now SF leads 6,5 - 5,5. I was expecting gain of 1-5 elo at this level .

[Jouni]

After 7 gamepairs 2 - 0 for Stockfish. With buggy version was 5 - 0.

[Jouni]

After 10 game pairs this is heading to 60-40 result. Quite similar to seasons 26-28 with LCZero. LCZero won game pairs in every final. Reckless not yet.