Finally official Lc0 v0.31.0-rc1

[Werewolf]

looks like the latest lc0 build has a few nasty bugs on 40xx cards. Hopefully there will be a fix soon.

[Jouni]

Hint. By renaming exe to lc0simple.exe You get rid of zillion uci parameters .

Code: Select all

       _
|   _ | |
|_ |_ |_| v0.31.0-rc1 built Mar 25 2024
uci
id name Lc0 v0.31.0-rc1
id author The LCZero Authors.
option name WeightsFile type string default <autodiscover>
option name Threads type spin default 0 min 0 max 128
option name MultiPV type spin default 1 min 1 max 500
option name Contempt type string default
option name SyzygyPath type string default
option name Ponder type check default true
option name UCI_Chess960 type check default false
option name UCI_ShowWDL type check default false
option name UCI_ShowMovesLeft type check default false
option name RamLimitMb type spin default 0 min 0 max 100000000
option name MoveOverheadMs type spin default 200 min 0 max 100000000
option name Slowmover type string default 1.000000
option name ValueOnly type check default false
option name LogFile type string default
uciok

[CornfedForever]

looks like the latest lc0 build has a few nasty bugs on 40xx cards. Hopefully there will be a fix soon.

Curious - what is happening that seems to be a 'bug'?

[Werewolf]

looks like the latest lc0 build has a few nasty bugs on 40xx cards. Hopefully there will be a fix soon.

Curious - what is happening that seems to be a 'bug'?

This is from this Discord (so I don't misrepresent):
"So, the v0.31 rc1 has a bug, where in the cuda-fp16 backend the BT4 network sometimes returns bad number.

It's not clear in which configurations it reproduces. So far we suspect that it's some combination of RTX40x0 card, new driver, Windows, and CUDA 11."

In addition in my testing I notice very variable nps (much more than usual, though this might have something to do with the new mini-batch settings) and evaluations fluctuating from +98% to 50% in obviously winning positions

[Dann Corbit]

Not all networks will work for me. The network used by TCEC does not run at all.

[Jouni]

Dann which GPUs you have?

[wickedpotus]

But what we DO KNOW, is that LC0 is faster and better on most modern cheap nvidia laptops, than on an úberpriced Apple computer

Yes and because of úberpriced Apple computer, everybody is buying those
Don't forget to mention that only some LC0 developers are tuning LC0 from the beginning for Nvidia GPUs, because they have only these GPUs and no AMD or Apple GPUs. So it's completely basic that LC0 is faster on Nvidia GPUs after all these years.
They could easily use some AMD GPUs and improve the speed of LC0 on those GPUs or on Intel GPUs.
The developers mentions years ago that they could easily improve the speed of LC0 by a factor of 5x but they are to much Nvidia guys.
= It's a developers / money problem and not an Apple, AMD or Intel problem.

First, let's address the elephant in the room: the notion that Apple's popularity is a testament to its superiority. While it's true that many people flock to Apple products, allow me to counter with a timeless adage: 'Eat shit because fifty billion flies can't be wrong.' Now, I'm not suggesting that Apple's products are akin to excrement, but rather, I'm highlighting the fallacy of equating popularity with superiority.

Now, let's talk about LC0 and its performance on various GPUs. It's true that some developers have focused their efforts on Nvidia, and why not? Nvidia has been the de facto standard and leader in AI hardware for quite some time. Their CUDA platform has been instrumental in advancing the field of deep learning, and it's only natural that developers would gravitate towards the most robust ecosystem.

As for the claim that LC0 could be five times faster with a little more effort, I'm reminded of the old adage: 'Talk is cheap. Show me the code.' It's easy to make grandiose claims, but it's quite another to back them up with tangible results.

In conclusion, while it's always entertaining to engage in a bit of playful ribbing, let's not lose sight of the facts. Nvidia's dominance in the AI hardware space is well-deserved, and it's only natural that developers would prioritize their platform. But who knows? Perhaps one day AMD or Apple will rise to the challenge and give Nvidia a run for its money. Until then, let's let the machines do the talking, shall we?

[Werewolf]

But what we DO KNOW, is that LC0 is faster and better on most modern cheap nvidia laptops, than on an úberpriced Apple computer

Yes and because of úberpriced Apple computer, everybody is buying those
Don't forget to mention that only some LC0 developers are tuning LC0 from the beginning for Nvidia GPUs, because they have only these GPUs and no AMD or Apple GPUs. So it's completely basic that LC0 is faster on Nvidia GPUs after all these years.
They could easily use some AMD GPUs and improve the speed of LC0 on those GPUs or on Intel GPUs.
The developers mentions years ago that they could easily improve the speed of LC0 by a factor of 5x but they are to much Nvidia guys.
= It's a developers / money problem and not an Apple, AMD or Intel problem.

First, let's address the elephant in the room: the notion that Apple's popularity is a testament to its superiority. While it's true that many people flock to Apple products, allow me to counter with a timeless adage: 'Eat shit because fifty billion flies can't be wrong.' Now, I'm not suggesting that Apple's products are akin to excrement, but rather, I'm highlighting the fallacy of equating popularity with superiority.

Now, let's talk about LC0 and its performance on various GPUs. It's true that some developers have focused their efforts on Nvidia, and why not? Nvidia has been the de facto standard and leader in AI hardware for quite some time. Their CUDA platform has been instrumental in advancing the field of deep learning, and it's only natural that developers would gravitate towards the most robust ecosystem.

As for the claim that LC0 could be five times faster with a little more effort, I'm reminded of the old adage: 'Talk is cheap. Show me the code.' It's easy to make grandiose claims, but it's quite another to back them up with tangible results.

In conclusion, while it's always entertaining to engage in a bit of playful ribbing, let's not lose sight of the facts. Nvidia's dominance in the AI hardware space is well-deserved, and it's only natural that developers would prioritize their platform. But who knows? Perhaps one day AMD or Apple will rise to the challenge and give Nvidia a run for its money. Until then, let's let the machines do the talking, shall we?

It's not just chess. I'm lucky enough to have access - via work - to an M2 Ultra and a PC with a 4090. Try doing video upscaling on both of those and the Mac gets eaten alive.

I keep saying it:
The Mac's performance per Watt is good.
The Mac's overall performance is OK.
The Mac's performance per $ is poor.

[Werewolf]

Can anyone shed light on the new Lc0 rumours...?

[Werewolf]

Forget it, it was an April Fool’s joke