How much Rtx4080 help lc0

[Werewolf]

Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti.

Incredible. And the 3060 ti is no slouch.

So the 4090 is roughly double the 3090?

[mehmet123]

Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti.

Incredible. And the 3060 ti is no slouch.

So the 4090 is roughly double the 3090?

Lc0 at RTX 4090 is ~%90 faster than Lc0 at RTX 3090. ~33500 nps for T78xxx at RTX 4090 and ~17500 nps for T78xxx at RTX 3090.

If the development of GPU's continues at this rate, Lc0 will be more advantageous than Stockfish if we talk about equal priced GPUs and Processors.

[Werewolf]

Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti.

Incredible. And the 3060 ti is no slouch.

So the 4090 is roughly double the 3090?

Lc0 at RTX 4090 is ~%90 faster than Lc0 at RTX 3090. ~33500 nps for T78xxx at RTX 4090 and ~17500 nps for T78xxx at RTX 3090.

If the development of GPU's continues at this rate, Lc0 will be more advantageous than Stockfish if we talk about equal priced GPUs and Processors.

Thanks.

And do you know how scaling works with multiple GPUs? Is 2 GPUs twice as fast as one, or is there an efficiency drop? I'm wondering both in terms of nps and search speed...

[mehmet123]

And do you know how scaling works with multiple GPUs? Is 2 GPUs twice as fast as one, or is there an efficiency drop? I'm wondering both in terms of nps and search speed...

Some Lc0 scaling tests made by Viren6
https://github.com/Viren6/lc0/wiki/Scal ... to-8x-A100

It would not be wrong to say that scaling in Lc0 differs according to GPU and Lc0 net.

[Werewolf]

And do you know how scaling works with multiple GPUs? Is 2 GPUs twice as fast as one, or is there an efficiency drop? I'm wondering both in terms of nps and search speed...

Some Lc0 scaling tests made by Viren6
https://github.com/Viren6/lc0/wiki/Scal ... to-8x-A100

It would not be wrong to say that scaling in Lc0 differs according to GPU and Lc0 net.

That's really interesting. The nps scaling from 1 to 2 GPUs almost seems too good.

Do you know if there's a reduction in search efficiency as there is with Alpha Beta engines going from 1 core to 2?
I don't know how multi-GPU efficiency works with Monte Carlo.