TCEC S29 division P started

[Damir]

It has nothing to do with hardware, but architecture. Lc0 is simply poorly designed, to perform, compared to Stockfish.

[Norm Pollock]

S29-divisionP was a lively and interesting tournament.

Some observations:

The 2 promoted engines from S29-League1, Reckless and Torch, finished #2 and #5.

The 2 promoted engines from S28-League1, PlentyChess and Integral, finished #4 and #6.

Perennial super-engines Stockfish and LcZero finished #1 and #3.

I don't understand how PlentyChess had 2 engine crashes early on, and then was fine.

The number of decisive games (wins) increased, indicating more bias in the openings. All in all, there were 91 wins in 224 games, but really 89 in 222 (40.1%) since there were 2 crashes by PlentyChess. And importantly, White had all the wins, Black had zero wins.

The finals should be great!

[Jouni]

Final is running now.

[Dann Corbit]

It has nothing to do with hardware, but architecture. Lc0 is simply poorly designed, to perform, compared to Stockfish.

If you are discussing millions of operations per Elo generated, then you are definitely correct.

On the other hand I know some people who have very high end GPUs and LC0 on those platforms is amazing, even tactically. My poor old 2080 Super GPUs can't begin to compete with those, even given a huge time headstart.

Once the memory sharing architecture used on the commercial GPUs becomes available on the consumer GPUs, we will see a new sort of power radiating from GPU powered chess programs.

[lucario6607]

It has nothing to do with hardware, but architecture. Lc0 is simply poorly designed, to perform, compared to Stockfish.

If you are discussing millions of operations per Elo generated, then you are definitely correct.

On the other hand I know some people who have very high end GPUs and LC0 on those platforms is amazing, even tactically. My poor old 2080 Super GPUs can't begin to compete with those, even given a huge time headstart.

Once the memory sharing architecture used on the commercial GPUs becomes available on the consumer GPUs, we will see a new sort of power radiating from GPU powered chess programs.

Memory sharing architecture?

[Dann Corbit]

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

[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

Hehe, I am meanwhile out of the GPGPU loop. Do not read the white papers anymore. There is AMD Infinity Fabric, there is Nvidia NVLink, there is CXL over PCIe. If you want to share the memory, you have to connect the processing units to it somehow. Apple offers impressive memory bandwith for CPUs, and it has a true unified memory architecture across CPU/GPU/NPU, now wonder if we consider that the M-series is basically a SoC.

--
Srdja

[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

Hehe, I am meanwhile out of the GPGPU loop. Do not read the white papers anymore. There is AMD Infinity Fabric, there is Nvidia NVLink, there is CXL over PCIe. If you want to share the memory, you have to connect the processing units to it somehow. Apple offers impressive memory bandwith for CPUs, and it has a true unified memory architecture across CPU/GPU/NPU, now wonder if we consider that the M-series is basically a SoC.

--
Srdja

Haha, I am already outdated, you were right, AMD and others are working on UALink and Ultra Ethernet, the former for scale up the latter for scale out for HPC/AI.

https://en.wikipedia.org/wiki/UALink
https://ualinkconsortium.org/
https://ultraethernet.org/

--
Srdja

[Werewolf]

There are plans already to update GPU server to 16xRTX 6090 to give LCZero still chance .

In 2027 or so then. That card is a long way off sadly.

[lucario6607]

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.