GPU Shortage Seems To Be Ending

[towforce]

There seems to be the beginning of a price war between AMD and Nvidia. Could finally be the beginning of the end of the shortage. These things seem to last forever when you live through them!

https://www.techradar.com/news/amd-vs-n ... y-dropping

[smatovic]

Meanwhile a single CPU core AB-NNUE reaches the performance of an high-end GPU MCTS-PUCT/CNN? What is Lc0 still good at? Maybe an heterogeneous search approach as Dann proposed is the future? Run different searches with different nets on CPU+GPU+TPU+FPGA+ASIC+....

--
Srdja

[towforce]

Meanwhile a single CPU core AB-NNUE reaches the performance of an high-end GPU MCTS-PUCT/CNN?

Sorry, have I missed something? Does the above relate to something that has actually happened or is trending to happen soon? If not, what is is about? Or is it just a quip?

[smatovic]

Meanwhile a single CPU core AB-NNUE reaches the performance of an high-end GPU MCTS-PUCT/CNN?

Sorry, have I missed something? Does the above relate to something that has actually happened or is trending to happen soon? If not, what is is about? Or is it just a quip?

https://computerchess.org.uk/ccrl/404/

SF 15 NNUE, 8 core: 3746 Elo

Lc0 0.28.0 64-bit w752187, RTX 2080: 3665 Elo

Elo gain for SF NNUE from 1 to 4|8 core estimated to be in ~30|30+ Elo range:

https://talkchess.com/forum3/viewtopic. ... 47#p925547

https://talkchess.com/forum3/viewtopic. ... 20#p925311

--
Srdja

[towforce]

Thanks for the helpful links!

[towforce]

What is Lc0 still good at?

An ideal chess engine would never get beaten. We know that search isn't going to do this: quick hand-written evals are nowhere near the current level now. Knowledge might be able to do it.

Humans sometimes mis-evaluate positions badly: this must happen, or they'd never lose a game.

Maybe this training on billions of positions isn't the way, though: we need to find a way to force NNs to find the deep patterns that could make all the difference - even as far as understanding tiny differences that make the difference in very similar positions.

How about... train the NN on a tiny number of positions, then use the billion position evaluated database in the following way:

* find a position from that billion that it evaluates badly

* add this position to its list of training games

* retrain the NN so that it evaluates all positions approximately correctly

* see if any games (apart from the latest) can be removed from the training list, the NN retrained, and it still evaluate everything it has so far correctly (including the removed game)

* repeat

My hope would be that this method would force the NN to find the deep patterns that it misses when it trains on a small number of positions, and that finding these would enable it to evaluate quickly and accurately with a much smaller NN (without a graphics card!).

[Jouni]

How about A100? TCEC uses 2xA100 with about 20 000 € price . Stockfish beats currently Lc0 badly. Slower GPUs have no chance against 8 core CPU.

[towforce]

How about A100? TCEC uses 2xA100 with about 20 000 € price . Stockfish beats currently Lc0 badly. Slower GPUs have no chance against 8 core CPU.

The A100 gets a mention in this article in the "petascale" section - link. I drool at the specifications - but I'm leaving it off my shopping list for now.

[ThomasMiller]

That's a good news but at this point I would wait for October when rumors say new GPU (Ampere Next/Ada Lovelace) could be released by Nvidia. New AMD GPU (RDNA 3) should come out within 2022 as well.

[smatovic]

...Intel Arc Alchemist discrete GPU is planed for H2/22 release, so three gamer gpu vendors in competition this year.

--
Srdja