Latest lc0-win-20180604-cuda92-cudnn714-experimental EXCELLENT

[Werewolf]

The latest lc0 CUDA experimental version is really excellent !
I'll be there, even if I have to buy a Titan V for that to happen !

I don’t think the Titan V is that much faster than your card without using its Tensor cores, maybe 30% or so?

Also given the attention on AI - and that Nvidia are in no hurry to make gaming cards - I suspect a new Titan may be out in 6-9 months.

Running 2x 1060 even 3GB version is at least 20% stronger than running 1080ti and 2x1060 3GB cost at least 30% less than single 1080Ti .

That's a possibility I didn't realise lc0 could exploit - is it hard to get it to run on 2 GPUs? Presumably if money was no object one could run 2 x 1080ti and get something faster than a Titan V (without its Tensor cores) ?

[shrapnel]

if you did 2 or 3 mins per move, do you think lc0 may have a chance to win ?

Almost certainly. If not now, within 1-2 weeks.
One problem is that even in 1 minute/move,it very rarely uses the full one minute, and makes its move within 25-30s on an average.

[shrapnel]

Running 2x 1060 even 3GB version is at least 20% stronger than running 1080ti and 2x1060 3GB cost at least 30% less than single 1080Ti .

I'd simply plonk in another 1080 Ti when Albert enables Multi-GPU support and still be ahead of you, didn't that occur to a smart guy like you, Milos ?

[Milos]

I don’t think the Titan V is that much faster than your card without using its Tensor cores, maybe 30% or so?

Also given the attention on AI - and that Nvidia are in no hurry to make gaming cards - I suspect a new Titan may be out in 6-9 months.

Running 2x 1060 even 3GB version is at least 20% stronger than running 1080ti and 2x1060 3GB cost at least 30% less than single 1080Ti .

That's a possibility I didn't realise lc0 could exploit - is it hard to get it to run on 2 GPUs? Presumably if money was no object one could run 2 x 1080ti and get something faster than a Titan V (without its Tensor cores) ?

It's fairly trivial. You just need multiplexing backend, i.e. something like:
--backend=multiplexing "--backend-opts=(backend=cudnn,gpu=0,threads=2),(backend=cudnn,gpu=1,threads=2)"
Btw. Titan V is already using Tensor cores, but the thing is that Tensor cores are not even remotely efficient as NVIDIA advertises them. In some workloads (mainly training) one can get 2.5x 1080Ti performance, in others like Lc0-cudnn inference it is only 1.3x.

[Ozymandias]

Interesting! That contradicts this report

https://www.pcgamer.com/nvidia-ceo-says ... orce-gpus/

I actually watched a video where he says "it'll be *gestures with his hand in a dismissive way* a long time away"

I guess they have leftovers to get rid of, from the crypto-bubble.

[Milos]

Running 2x 1060 even 3GB version is at least 20% stronger than running 1080ti and 2x1060 3GB cost at least 30% less than single 1080Ti .

I'd simply plonk in another 1080 Ti when Albert enables Multi-GPU support and still be ahead of you, didn't that occur to a smart guy like you, Milos ?

Your trollish behaviour is really stepping over the line. So you are going to my ignore list.
This is not pissing context and who has a bigger one. It seems Indian ppl are really full of complexes related to material stuff. I use quite old GTX770 card and it is perfectly fine for my needs. If I need (for work) I have access to the cluster with 10 TitanV's and 50 1080Ti's. And if I wanted to I could buy 20 1080Ti's just from monthly salary since [moderation: removed remark insulting for Indians in general]

[Werewolf]

Running 2x 1060 even 3GB version is at least 20% stronger than running 1080ti and 2x1060 3GB cost at least 30% less than single 1080Ti .

That's a possibility I didn't realise lc0 could exploit - is it hard to get it to run on 2 GPUs? Presumably if money was no object one could run 2 x 1080ti and get something faster than a Titan V (without its Tensor cores) ?

It's fairly trivial. You just need multiplexing backend, i.e. something like:
--backend=multiplexing "--backend-opts=(backend=cudnn,gpu=0,threads=2),(backend=cudnn,gpu=1,threads=2)"
Btw. Titan V is already using Tensor cores, but the thing is that Tensor cores are not even remotely efficient as NVIDIA advertises them. In some workloads (mainly training) one can get 2.5x 1080Ti performance, in others like Lc0-cudnn inference it is only 1.3x.

Brilliant thanks.

[Albert Silver]

Interesting! That contradicts this report

https://www.pcgamer.com/nvidia-ceo-says ... orce-gpus/

I actually watched a video where he says "it'll be *gestures with his hand in a dismissive way* a long time away"

I guess they have leftovers to get rid of, from the crypto-bubble.

I doubt it. They refused to ramp up the production to meet the demand, which is what led to the absurd price gouging many retailers began charging.

[Werewolf]

I'm trying to understand the improvement lc0 is making when the weights here

http://lczero.org/networks

don't seem to be making any progress at all. Is it that the network is saturated but other settings to do with CUDA still offer room for improvement?

[whereagles]

a quick linear regression from ID323 to ID379 shows a 0.551 average ELO gain per ID (p-value ~0%, but R^2 = 0.35 only...)

evidence hints at some climbing still going on, but a lot of noise as well.