I stumbled upon this article on the new Nvidia RTX GPUs

[Laskos]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

I didn’t stay for the whole thing but the gist of it was:

1) Lc0 uses tensor cores
2) new tensor performance has 2 values
3) Lc0 uses the lower one, which means the 3080 will be about the same as a 2080Ti.

We need clarity over premise 1 for starters.

I see, thanks. I see these values here for example, but I don't understand the issue:

https://www.anandtech.com/show/16057/nv ... 0-rtx-3090

Would be nasty if power consumption increases as much as the performance.

[Laskos]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

If you are able to run 2 cards in your system, and it is for running Lc0. Buying a RTX 3070 for $499 would be great. RTX 2070+RTX 3070, and this should give you some very nice gains when running Lc0.

After 3070 appears and it disappoints us, 2070 and 2070 Super will go on cheapo, and I might go for additional one of those.

[Werewolf]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

I didn’t stay for the whole thing but the gist of it was:

1) Lc0 uses tensor cores
2) new tensor performance has 2 values
3) Lc0 uses the lower one, which means the 3080 will be about the same as a 2080Ti.

We need clarity over premise 1 for starters.

I see, thanks. I see these values here for example, but I don't understand the issue:

https://www.anandtech.com/show/16057/nv ... 0-rtx-3090

Would be nasty if power consumption increases as much as the performance.

So one issue is that although the new tensor cores are better than the old ones, there are fewer of them in the new cards.

Anyway, here's an interesting read

https://en.wikipedia.org/wiki/List_of_N ... _30_series

scroll right until you get to tensor TFLOPS and you have to ignore sparse because apparently Lc0 doesn't use it.
From this alone the figures look ok, so I'm not sure what the issue was.

[Lars Sandin]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

If you are able to run 2 cards in your system, and it is for running Lc0. Buying a RTX 3070 for $499 would be great. RTX 2070+RTX 3070, and this should give you some very nice gains when running Lc0.

Unfortunately it doesn’t seem to be quite as simple this time around.

https://www.tomshardware.com/news/nvi ... r-rtx-3090

[mwyoung]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

If you are able to run 2 cards in your system, and it is for running Lc0. Buying a RTX 3070 for $499 would be great. RTX 2070+RTX 3070, and this should give you some very nice gains when running Lc0.

Unfortunately it doesn’t seem to be quite as simple this time around.

https://www.tomshardware.com/news/nvi ... r-rtx-3090

I do not think Lc0 is using this tech when running Lc0 on 2 or more cards. Does anyone know for sure.

[JohnW]

Here are some gaming benchmarks..

https://www.dsogaming.com/news/first-le ... enchmarks/

[Laskos]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

I didn’t stay for the whole thing but the gist of it was:

1) Lc0 uses tensor cores
2) new tensor performance has 2 values
3) Lc0 uses the lower one, which means the 3080 will be about the same as a 2080Ti.

We need clarity over premise 1 for starters.

I see, thanks. I see these values here for example, but I don't understand the issue:

https://www.anandtech.com/show/16057/nv ... 0-rtx-3090

Would be nasty if power consumption increases as much as the performance.

So one issue is that although the new tensor cores are better than the old ones, there are fewer of them in the new cards.

Anyway, here's an interesting read

https://en.wikipedia.org/wiki/List_of_N ... _30_series

scroll right until you get to tensor TFLOPS and you have to ignore sparse because apparently Lc0 doesn't use it.
From this alone the figures look ok, so I'm not sure what the issue was.

Anandtech and Wiki seem to disagree exactly in our case. I misread something?

[Werewolf]

The Anandtech article does indeed use different figures. Very odd.

My interpretation of the Anandtech article is negative though, since it claims 2080Ti tensor as 114; and 3080 as 119 TFLOPS (note fp16 non sparse).

That’s a tiny improvement.

[Milos]

https://www.engadget.com/nvidia-rtx-309 ... 59544.html

If I understood correctly, the improvement in our case compared to 2xxx generation will be great, right?
If yes, I am starting putting aside bit by bit, so that in November I will get a 3070.

The real deal will be when (if ever) GPU and CPU functionality are on one chip, able to seamlessly address the same memory and there are included asm instructions to perform matrix operations (plus C+++)

That will never happen at least for high-end GPUs, mostly due to thermal problems/packaging of the SoC that would contain both CPU and GPU. What we really need is GPU that can access main memory seamlessly and concurrently with CPU.

[Milos]

A very long conversation on Lc0 discord yesterday suggested they wouldn't be that great

The reason? Not as great as possible 150% improvement or much much lower than that?

viewtopic.php?f=2&t=72320&p=846639&#p846617