Scaling of Lc0

[Hugo]

Me again

I am quite new with Lc0, but highly interested.
What could be the ELO benefit from 1 x RTX 2070 to 2 x RTX 2070 ?
In analysis I noticed that the main line does not change after 10 - 20 seconds.
So maybe the benefit is very low compared to energy hunger and the investment of a second RTX card.

C.K.

[MikeB]

Me again

I am quite new with Lc0, but highly interested.
What could be the ELO benefit from 1 x RTX 2070 to 2 x RTX 2070 ?
In analysis I noticed that the main line does not change after 10 - 20 seconds.
So maybe the benefit is very low compared to energy hunger and the investment of a second RTX card.

C.K.

It's not 100% clear at this point - but typically doubling of speed by adding a second CPU would be capped to 50 or 60 Elo. Similar to a CPU, there is no evidence, to my knowledge, that second GPU would offer more than a 60 Elo gain and perhaps the gain would be less than 60 Elo. If true, the return is very low when compared to the energy cost and the investment of a second RTX card, especially as compared to the return of the first GPU and its cost. I welcome any additional comments from anyone who has actually experimented with this.

[Hugo]

The main question to me is, how does Leela play chess?
AB engines calculate, and the more cpu power they have, the more can be calculated (theoretical)

Leela seems to be different. Maybe I am wrong, but it seems like Leela is searching in her weights file.
BAsed on what can be found there, the actually board possition will be estimated.
I have not seen yet, that a variant has changed after 2 or more minutes. Thats why I doubt that doubling the GPU power has same effect as on AB engines.

C.K.

[henk2]

The main question to me is, how does Leela play chess?
AB engines calculate, and the more cpu power they have, the more can be calculated (theoretical)

Leela seems to be different. Maybe I am wrong, but it seems like Leela is searching in her weights file.
BAsed on what can be found there, the actually board possition will be estimated.
I have not seen yet, that a variant has changed after 2 or more minutes. Thats why I doubt that doubling the GPU power has same effect as on AB engines.

C.K.

Leela uses PUCT

http://citeseerx.ist.psu.edu/viewdoc/do ... 1&type=pdf

Which is very similar to MCTS.

"The focus of Monte Carlo tree search is on the analysis of the most promising moves, expanding the search tree based on random sampling of the search space. The application of Monte Carlo tree search in games is based on many playouts. In each playout, the game is played out to the very end by selecting moves at random. The final game result of each playout is then used to weight the nodes in the game tree so that better nodes are more likely to be chosen in future playouts. "

But Lc0 doesn't do game rollouts. So it's a little different from a pure MCTS.

Your doubt that doubling GPU power doesn't have the same effect as on an AB engine is correct.
But GPU power is still increasing at an exponential rate while CPU power (especially single core performance) has stopped following exponential growth.
But now that AMD is finally competitive with Intel again we might see faster gains again.

[henk2]

Me again

I am quite new with Lc0, but highly interested.
What could be the ELO benefit from 1 x RTX 2070 to 2 x RTX 2070 ?
In analysis I noticed that the main line does not change after 10 - 20 seconds.
So maybe the benefit is very low compared to energy hunger and the investment of a second RTX card.

C.K.

It's not 100% clear at this point - but typically doubling of speed by adding a second CPU would be capped to 50 or 60 Elo. Similar to a CPU, there is no evidence, to my knowledge, that second GPU would offer more than a 60 Elo gain and perhaps the gain would be less than 60 Elo. If true, the return is very low when compared to the energy cost and the investment of a second RTX card, especially as compared to the return of the first GPU and its cost. I welcome any additional comments from anyone who has actually experimented with this.

Since Leela's search is very similar to MCTS, it converges to minimax slower/less efficiently than an A/B search would.
So MCTS may miss a move and never find it, it's especially evident in late game where Leela often "blunders."

But GPU power is increasing at a much faster pace than CPU power so she does have an inherent advantage.
She may not see the same increase in ELO per doubling of processing power, but her hardware will double in power in much less time than the hardware of a CPU engine.
Although Leela's fairly strong on just a CPU too. Still about 200-250 ELO removed from Ethereal.

https://github.com/dkappe/leela-chess-w ... d-Networks

[Laskos]

Your doubt that doubling GPU power doesn't have the same effect as on an AB engine is correct.

I don't quite understand that statement. On single GPU, Leela does scale well with NPS or nodes searched in the same amount of time. One can tune things like CPuct and similar search factors that it scales similarly to SF at both short-TC and LTC. TCEC shows that by choosing proper tuning, Leela scales well to LTC. So, the issue is scaling on multiple GPUs, not better single GPU. I am quite confused now about the scaling to 2 GPUs, but I would guess that the scaling is good. 4 GPUs is another matter.

[Werewolf]

Your doubt that doubling GPU power doesn't have the same effect as on an AB engine is correct.

I also don't understand this comment. Going from 1 GPU to 2 should yield around 50 elo.