TCEC S15, END of an ERA event is much more Brutal than I thought!

[Nay Lin Tun]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.

cap4.PNG

[Laskos]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

Yes, I expected long ago 2:1 Win/Loss ratio for Leela, something on the lines +14 -7 =79. Let's see. I predicted that when the previous S14 was underway with 32930 net.

It IS an end of an era, if anything, the hardware for Leela is below the hardware for Stockfish in the current TCEC. GPU has much more room for improvement than CPU, we will assist in 1-2 years at a complete dominance strength-wise of NN-bases engines. FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

[Werewolf]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

Not exactly fair for something like TCEC, but I think Stockfish's hardware saviour is a cluster. If a decent cluster was built with 1000+ cores there may be no version of Lc0 on any machine which could hold its own.

[George Tsavdaris]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

Not exactly fair for something like TCEC, but I think Stockfish's hardware saviour is a cluster. If a decent cluster was built with 1000+ cores there may be no version of Lc0 on any machine which could hold its own.

But why you assume no diminishing returns of real strength with 1000 cores? Most probably 1000 cores compared to 43 as of now on TCEC Stockfish would not give a gigantic Elo boost but only small one.
They tested 384 threads(on 192 cores) vs 32 threads(cores) for Stockfish and it was +114 Elo.

Even worse, even MUCH worse i would say, 1000 cores system is a >300 000 $ one while current TCEC Leela's GPUs is just 2000 $.

[Nay Lin Tun]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

Not exactly fair for something like TCEC, but I think Stockfish's hardware saviour is a cluster. If a decent cluster was built with 1000+ cores there may be no version of Lc0 on any machine which could hold its own.

Are you aware that 1000 cores= 22x44 cores only. That means it is about equivalent to 10x times Time control boost (because of SMP scaling whereas 2x TC is approx equal to 4x hardware especially beyond 32 cores).
You will be surprised that Leela actually already beat 44 cores (SF with 10x TC bonus ) in one pair of opening! 3 mins vs 30 mins.

[Laskos]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

Not exactly fair for something like TCEC, but I think Stockfish's hardware saviour is a cluster. If a decent cluster was built with 1000+ cores there may be no version of Lc0 on any machine which could hold its own.

But why you assume no diminishing returns of real strength with 1000 cores? Most probably 1000 cores compared to 43 as of now on TCEC Stockfish would not give a gigantic Elo boost but only small one.
They tested 384 threads(on 192 cores) vs 32 threads(cores) for Stockfish and it was +114 Elo.

Even worse, even MUCH worse i would say, 1000 cores system is a >300 000 $ one while current TCEC Leela's GPUs is just 2000 $.

This thread of Peter Osterlund is helpful with Lazy SMP and clusters:
viewtopic.php?f=7&t=64824

The cluster latencies prevent high gain from doubling the number of nodes in the cluster. I guess an 1,000 core cluster will barely be 3x-4x times better in "effective speed" compared to the TCEC machine. And this will soon be not enough to beat T40 newer and improving nets even on only 2-4 GPUs. Maybe it's not enough already. Also, GPUs will improve solidly soon, and NN engines' prallelization to 4 or 8 GPUs too.

Interesting times, traditional engines will become more like a niche "puzzle solvers".

[Leo]

SF has a chance if LZ throughs away some games like it did last time.

[glennsamuel32]

You will be surprised that Leela actually already beat 44 cores (SF with 10x TC bonus ) in one pair of opening! 3 mins vs 30 mins.

No point having a million cores searching the wrong lines.

[Hai]

After 26 games in TCEC SUFI, Leela is leading with 15-11.

My prediction for TCEC was just 55-45 Leela win.
cap4.PNG

Yes, I expected long ago 2:1 Win/Loss ratio for Leela, something on the lines +14 -7 =79. Let's see. I predicted that when the previous S14 was underway with 32930 net.

It IS an end of an era, if anything, the hardware for Leela is below the hardware for Stockfish in the current TCEC. GPU has much more room for improvement than CPU, we will assist in 1-2 years at a complete dominance strength-wise of NN-bases engines. FPGA might come next for chess engines, but it's a bit more involved to make them play chess.

LC0 vs Stockfish will end: +16 =80 -4

It was a big mistake to take Stockfish and not Bluefish!

[JJJ]

Stockfish is now the Komodo of lczero. It is first in rating list but loosing in direct confontation.
End of Era will be when Lczero will be the first in rating list as well.