CCC has serious hardware update!

[Nay Lin Tun]

256 cores is more than many companies had in the 1990s, maybe the whole universities.

It is 128 cores ( 256 threads). OP should have said 256 threads.

[shinkarom]

Even 128 cores is a city in itself.

[mwyoung]

CPUs: 2 x AMD EPYC 7H12
GPU: 2x A100 (40 GB GPU memory)
Cores: 256 cores (128 physical)
RAM: 512GB DIMM DDR4 2933 MHz (0.3 ns)
SSD: 2x Micron 5210 MTFD (2TB) in RAID1
OS: CentOS 8

What's TCEC answer ?

Why do they need to answer. Both have massive hardware. Will it change any results.

If you notice even on our much smaller hardware. The ranking of the engines does not really change.

And this is also what I am seeing in my testing.

Yes, I don't see the point in this huge hardware. As you say, the results are probably the same on 32 cores, and I'd guess none of the authors have tested their engines on this level of hardware. Better quality chess however you define that, maybe, but you also probably increase the draw rate as well. Which is when they have to move towards unbalanced openings as an artificial way to spice things up.

Yes, and more cores give you less and less real speed. Because of diminishing returns. As you can only effectively subdivide the search of the engine so many times. What is a bigger benefit in the new hardware. Is the improved IPC and frequency of each core.

[Nay Lin Tun]

It is ture.
In previous HCE of Stockfish, 250 million nodes / position search is the sweet spot of Stockfish. Beyond that, elo dimish significantly. ( This could be true to SFNNUE )

There could be +20 elo from 250 million nodes to 1 billion nodes.

And there could be +20 elo from 1 billion nodes to 4-5 billion nodes.

Doubling hardware from 64 cores to 128 cores is expected to gain 10-15 elo only.

Here is nodes vs elo test.

[mwyoung]

It is ture.
In previous HCE of Stockfish, 250 million nodes / position search is the sweet spot of Stockfish. Beyond that, elo dimish significantly. ( This could be true to SFNNUE )

There could be +20 elo from 250 million nodes to 1 billion nodes.

And there could be +20 elo from 1 billion nodes to 4-5 billion nodes.

Doubling hardware from 64 cores to 128 cores is expected to gain 10-15 elo only.

Here is nodes vs elo test.

We are bursting many bubbles here today. And I see by the dislikes on my stream proving this point is not making some happy.

What does a 10x NPS advantage gain you. So far nothing after 7 games...

[Joost Buijs]

The point is that it is all bull sh*t, with so many threads on slow CPU's the engines will only degrade.

[Dann Corbit]

I think that is the main value of contests like CCC and TCEC.
They drive innovation.
In the early TCEC contests, engines crashed and burned like strings of chinese fire crackers.
So the engine writers were forced to write robust code or bow out (there was one contest where they had to invite Crafty, because there were no other well known chess engines that would scale to TCEC hardware).
Struggling to get SMP to work well is what lead to LazySMP.
Maybe there is a new miracle just around the corner.

[mwyoung]

I think that is the main value of contests like CCC and TCEC.
They drive innovation.
In the early TCEC contests, engines crashed and burned like strings of chinese fire crackers.
So the engine writers were forced to write robust code or bow out (there was one contest where they had to invite Crafty, because there were no other well known chess engines that would scale to TCEC hardware).
Struggling to get SMP to work well is what lead to LazySMP.
Maybe there is a new miracle just around the corner.

If they were smart. What they can do is use these wasted cores. In running more then one tournament, or sets of games at a time. That is the real advantage in the massive hardware right now. Is in faster and also accurate results.

[Alayan]

The new CCC hardware has both more cores and faster cores than before. And the GPU are definitely proving faster than 4xRTX2080Ti did for Leela.

Both CPU and GPU engines are seeing a significant speed uplift, though of course there are always diminishing returns on more speed.

Look at the specs for the max extra clock speeds...they are very low for the CPUs with a huge number of cores.

Seems like the optimal LC0 server might have 8 cores at 7 or 8 GHz cooled with liquid nitrogen for the CPU part.
It would probably still cost less than what they are using.

Liquid nitrogen is completely insane for production use. Nobody ever does that for stable use, liquid nitrogen is only for overclocking records.

[AndrewGrant]

Finally! Good riddance to the over priced, over used, under clocked, under performing Intel garbage.

CCC's cores are faster than TCEC.
CCC's setup now has more cores than TCEC.
CCC's machines have better hyperthreading.

CCC is now producing the highest quality chess games available anywhere to the public.