Best Hardware for Leela (Lc0 Alpha Zero) in 2023

https://lczero.org/dev/wiki/best-nets-for-lc0/

How do you build these nets from the gpu of your video card ?
How should someone who is not a tech guide can set the correct parameters (I've been tried googling it a lot and I cannot understand)

What is the recommended GPU for Leela to make it more accurate for high level anyalysis (I am 2350 FIDE FM) - I intend to use LEELA a lot (not only waste money on Ducats on Chessbase Server or use CHessify server) to create high quality Chessable Courses.
Is there a difference in terms of quality of analysis (tactical accuracy of RTX 3080 TI to 4080 TI or 4090 TI ) does it make any difference to Leela ?

I suggest asking on talkchess.com, where many engine developers discuss computer chess related matters, including hardware. You will get the best advice there.

I think it may not be a good idea to build your own nets (unless you want to get into computer chess), but instead find out what net is best for your hardware, and installing it. It is easy to install different nets and even run a tournament between them. Once I did that and found that on my humble laptop a much smaller net was outperforming (winning more games against) a larger net.

As far as I know, alphazero is not available and most likely unmaintained at a private company. Lc0 started by implementing in opensource the alphazero algorithms as they were published. Lc0 developed a lot over the years. So most likely Lc0 of today will be significantly stronger than alphazero. In the last 6 TCEC seasons Stockfish won the final - 5 times against Lc0 and once against Komodo Dragon.

Correct, Alpha Zero is not available and never has been, and would be extremely outclassed by any high-end modern engine. Lc0 will run with superhuman strength on any modern NVidia GPU. On a 4090 (there is no 4090 Ti), I get about 50knps on startpos using net 785469 (512 filters, 20 blocks), which is a bit under double what I would expect from a 3080 Ti. The best hardware for Lc0 is probably a cluster of H100s. TCEC (the Top Chess Engine Championship) runs Lc0 on 2x A100 GPUs (plus 52 CPU cores, which isn't as relevant) and gets about the same 50knps speed with a 512x20 net.

Why do you get the same knps on a rig that costs like 5% of the TCEC one? Is knps not that important?
@corvusmellori

@Rookitiki said in #4:

Why do you get the same knps on a rig that costs like 5% of the TCEC one? Is knps not that important?
@corvusmellori

Yeah, I was surprised. There are a few differences. The first is that TCEC was using a newer build of Lc0 which gets slightly slower speeds. I installed that and my knps dropped by about 10%. But it's also just the case that the 4090 is a very, very good card. It will go faster than a single A100 and stay in the same ballpark as TCEC's dual A100.

Going from 3080ti to 4080ti gives a significant improvement in analysis. But going from 4080ti to 4090 gives only a marginal improvement (not noticeable in a vast majority of positions). 4090ti doesn't exist yet until later this year.

@corvusmellori said in #5:

Yeah, I was surprised. There are a few differences. The first is that TCEC was using a newer build of Lc0 which gets slightly slower speeds. I installed that and my knps dropped by about 10%. But it's also just the case that the 4090 is a very, very good card. It will go faster than a single A100 and stay in the same ballpark as TCEC's dual A100.
Also A100s are very old tech now. That's like comparing older 128 core xeon servers to a modern i9-13000k processer

@Lcolem said in #7:

Also A100s are very old tech now. That's like comparing older 128 core xeon servers to a modern i9-13000k processer
I've been considering putting together a dual EPYC 7763 system for Stockfish, which I expect to outperform my 7950x based system by at least 3x (admittedly with 6x the cost or so).