Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

[mehmet123]

Stockfish Cluster support at least 4096 cores. At this hardware the speed of Stockfish is ~4.800.000.00 kn/s.
Noobpwnftw (main contributor of Fishtest ) has a resource with ~30.000 cpu.
At 5000 TPU AlphaZero has played 44 millions self play games at 9 hours. And the speed of AlphaZero was 80 kn/s at 4 TPU GPU. At 2017 match AlphaZero (4 TPU) has beaten Stockfish 8 (64 core).

Which engine should we choose to represent the world in the event of a chess match with an alien race? AlphaZero at 5000 TPU or Stockfish 11 (Dev.) at 30.000 cores.

http://ipmanchess.yolasite.com/amd---in ... -bench.php

[CMCanavessi]

Leela is stronger than both.

[mehmet123]

Leela hasn' t a resource like AlphaZero. 80 millions self games have been playing by Lc0 for 9 months.

[Leo]

AlphaZero at 5000 TPU. Stockfish cant utilize all those cores.

[Laskos]

AlphaZero at 5000 TPU. Stockfish cant utilize all those cores.

AlphaZero doesn't scale that great either beyond 48 or so TPUs, if I remember the paper correctly.

[mehmet123]

AlphaZero at 5000 TPU. Stockfish cant utilize all those cores.

The speed of Stockfish at 4096 cores is ~4.800.000.000/second. It'a very exciting speed. I don't think Stockfish has a serious scaling problem.

If Stockfish has ability of to use 4096 cores so I don't think Stockfish will has a problem using 30.000 cores.

[Raphexon]

It doesn't work like that.

A/B doesn't scale that well.
30.000 cores would also need a different implementation than Lazy-SMP.

The real speedup might only be 50-100 for SF. (That's generous)

[mehmet123]

Stockfish (4 core) has 3493 elo according to CCRL (40/15). Stockfish 11 dev has ~3500 elo.
Stockfish speed is 4.800.000kn/s at 4096 cores. It's nearly 2000x of Stockfish at 4 cores. The speed of 2x Athlon 4600(main engine of CCRL then they changed the main engine but the elo of engines didn't change only time control was changed) is ~ 2500 kn/s

If we give 40 elo to doubling than Stockfish has 3500 + (11 x 40)=3940 elo. But I think to double the speed gains 30 elo (average) at this high kn/s. So Stockfish 4096 cores has 3500 + (11 x 30)= 3830. And if Stockfish gains 70 elo when the number of cores are 30.000 than the elo of Stockfish is 3900 elo.

My elo estimate for Stockfish at 30.000 cores is 3850 or 3900.

[Dann Corbit]

He was quoting a real figure with the 4 billion NPS.
http://ipmanchess.yolasite.com/amd---in ... -bench.php

Look at the cluster figure by vondele.

I know Amdahl's law must come into play at some point, and maybe the useful work increase does not scale with NPS, but it appears that LazySMP scales like a screaming son of a gun. (Hmm... is that good or bad?)

On the other hand, It may not be currently possible to physically scale a hardware system to the limits stated, and imagine the cost of the Stockfish cluster used in the real experiment.

And since you cannot buy a TPU, only Google could produce the TPU version. So we are once again left with the alternative form of the question,
"How many angels can dance on the head of a pin?"

[Raphexon]

Stockfish (4 core) has 3493 elo according to CCRL (40/15). Stockfish 11 dev has ~3500 elo.
Stockfish speed is 4.800.000kn/s at 4096 cores. It's nearly 2000x of Stockfish at 4 cores. The speed of 2x Athlon 4600(main engine of CCRL then they changed the main engine but the elo of engines didn't change only time control was changed) is ~ 2500 kn/s

If we give 40 elo to doubling than Stockfish has 3500 + (11 x 40)=3940 elo. But I think to double the speed gains 30 elo (average) at this high kn/s. So Stockfish 4096 cores has 3500 + (11 x 30)= 3830. And if Stockfish gains 70 elo when the number of cores are 30.000 than the elo of Stockfish is 3900 elo.

My elo estimate for Stockfish at 30.000 cores is 3850 or 3900.

It doesn't work like that.
Strength doesn't scale linearly with cores, NPS sort of does because LazySMP can keep widening the search. But as the search widens it also becomes less efficient.

SF 1 million nodes per move is weaker on 16 cores than 1, so 16xTC is stronger than 16 cores.

SF with 30k time odds would crush the 30.000 cores.
SF with 100x time odds likely easily beats a 30k core cluster.

Also LazySMP won't work properly on such a big cluster.
Cluster systems have a cool amount of NPS but their strength is not quite as good as their NPS would make you believe.


A0 would also experience scaling issues.