Stockfish 18 for STC Rating List

[Rebel]

The last 6 TCEC finals were between SF and Leela, SF won them all.

Who had the better hardware ?

This page give clues.

Snippets

Number of Cores / Threads for Stockfish and all others
Each engine can use up to 512 threads of the processors

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

Are there any hardware rants in favor for Leela ?

Let's discuss.

So this was one of my reasons to include the neglected second best engine after Stockfish = Leela in my testing.

[Vinvin]

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

8 GPU RTX 5090 means 8x21760 cores = 174080 cores for Leela.
It's not fair ! There's way more cores for Leela !

[Rebel]

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

8 GPU RTX 5090 means 8x21760 cores = 174080 cores for Leela.
It's not fair ! There's way more cores for Leela !

Type "gpu cores vs cpu cores" into Google.

[Vinvin]

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

8 GPU RTX 5090 means 8x21760 cores = 174080 cores for Leela.
It's not fair ! There's way more cores for Leela !

Type "gpu cores vs cpu cores" into Google.

Why ? I simply use your own argumentation to prove you are wrong.

[Rebel]

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

8 GPU RTX 5090 means 8x21760 cores = 174080 cores for Leela.
It's not fair ! There's way more cores for Leela !

Type "gpu cores vs cpu cores" into Google.

Why ? I simply use your own argumentation to prove you are wrong.

Well, cpu cores are not comparable with gpu cores, do some reading as I suggested.

Type "gpu cores vs cpu cores" into Google, gives :


CPU cores
are designed for sequential, complex, and varied tasks with low latency (high speed, few cores), acting as the system's brain for OS operations.

In contrast, GPU cores are numerous (thousands) but specialized, designed for high-throughput, parallel, and repetitive data tasks like graphics rendering and AI.

This video explains the key differences between CPUs and GPUs, including their core architecture:
Thumbnail van gerelateerde video
1m
TechPrep
YouTube • 4 mei 2024



Key Differences Summary

Core Count & Power: CPUs have a few (4-64+) powerful cores; GPUs have thousands of smaller, less powerful cores.
Processing Type: CPUs handle serial (one by one) processing; GPUs excel at parallel (many at once) processing.
Best Use Case: CPUs are ideal for general computing, gaming logic, and operating system tasks. GPUs excel at video rendering, machine learning, and simulations.
Architecture: CPU cores have large caches and complex control logic, whereas GPU cores prioritize SIMD (Single Instruction, Multiple Data) throughput.

When to Use Which

CPU: Necessary for responsiveness, complex decision-making, and tasks that cannot be easily broken into smaller, parallel, independent tasks.

GPU: Superior for tasks that can be broken into thousands of smaller tasks, such as matrix calculations in AI, video rendering, or high-end graphics.

Modern systems use them together, with the CPU managing operations and sending parallel data tasks to the GPU for accelerated processing.

[jorose]

Are you trying to argue TCEC hardware is CPU favored? Because SF won the last few Super Finals against Leela? Or what is the argument?

In Chess.com Rapid that recently ended, Leela convincingly lost the fight for fourth place to Plentychess. I think if you ask the developers of both SF and Leela they will have a consensus that SF is currently stronger than Leela, and would be expected to win a match under reasonable hardware conditions. I don't think there is currently a lot of evidence pointing to Leela being close to SF.

The TCEC system CPUs are over 2.5 years old, I can't even find them to purchase new. The refurbished ones go for around 5k USD. The CPUs have a combined TDP of 720 watts. Meanwhile, 5090 is a current gen GPU with market rate around 2.5k per GPU. The GPUs have a combined TDP of 4600 watts. None of this is even considering Leela also has access to 2x AMD EPYC 9175F.

I think it is an apples to oranges comparison, but by any metric I can think of, TCEC is not CPU favored.

[RubiChess]

TCEC GPU server starting from S28
CPU: 2x AMD EPYC 9175F (32 cores/64 threads)
GPU: 8x NVIDIA GeForce RTX 5090 32607MiB

CPU 512 threads versus 8 x GPU

Is that fair ??????????

8 GPU RTX 5090 means 8x21760 cores = 174080 cores for Leela.
It's not fair ! There's way more cores for Leela !

Type "gpu cores vs cpu cores" into Google.

Type 174080 / 512 in your calculator.
Or look for the money for purchasing the 2 CPUs and 8 GPUs.
And don't forget: You still need to purchase a CPU for Leela after purchasing the GPUs.

The last 6 TCEC finals were between SF and Leela, SF won them all.

Who had the better hardware ?

I lost to Usain Bolt last week. He must have had the better shoes.

[AndrewGrant]

Are you trying to argue TCEC hardware is CPU favored? Because SF won the last few Super Finals against Leela? Or what is the argument?

In Chess.com Rapid that recently ended, Leela convincingly lost the fight for fourth place to Plentychess. I think if you ask the developers of both SF and Leela they will have a consensus that SF is currently stronger than Leela, and would be expected to win a match under reasonable hardware conditions. I don't think there is currently a lot of evidence pointing to Leela being close to SF.

The TCEC system CPUs are over 2.5 years old, I can't even find them to purchase new. The refurbished ones go for around 5k USD. The CPUs have a combined TDP of 720 watts. Meanwhile, 5090 is a current gen GPU with market rate around 2.5k per GPU. The GPUs have a combined TDP of 4600 watts. None of this is even considering Leela also has access to 2x AMD EPYC 9175F.

I think it is an apples to oranges comparison, but by any metric I can think of, TCEC is not CPU favored.

It is hard to come up with a metric by which TCEC can be considered CPU favoured.
In fact, none come to mind immediately. $ values, and wattage, are some of the more obvious ones, and those are very lopsided.

CCC's hardware has been stable for many years now, with the 2x EPYC 7H12 vs 2x A100 setup. And since it has been stable for many years, we can track the progress of Lc0, Stockfish, and others, in terms of software accomplishments not hardware ones. It has been the case that Leela has fallen from contention from 1st, down to at 5th place in the Rapid (10+3) events, and even lower in Bullet/Blitz events.

[Modern Times]

My amateur opinion - in TCEC, the hardware advantage feels massively in Lc0 favour to me.

That is on the assumption that Lc0 in fact performs better on 8x NVIDIA GeForce RTX 5090 than it would on say 4 or even 2. What do we know in that regard ?