Stockfish One Thread Challenge

[Jouni]

In CCC one thread SF (NNUE) beats 250 threads Ethereal! And plays equal to 250 threads Komodo or SF classic. Definitely doubling of cores don't give 70 ELO anymore .

[dkappe]

In CCC one thread SF (NNUE) beats 250 threads Ethereal! And plays equal to 250 threads Komodo or SF classic. Definitely doubling of cores don't give 70 ELO anymore .

Looking at the first event, Dragon had 37 point out of 60 vs 1 thread SF. Now I don’t know if that’s a better or worse than expected result, but it’s definitely not “equal.”

[mar]

Definitely doubling of cores don't give 70 ELO anymore .

doubling of time, not doubling of cores. no perfect scaling

[Jouni]

Currently SF - Komodo 15,5-16,5. Quite equal. I wasn't saying anything about Dragon.

[lkaufman]

Definitely doubling of cores don't give 70 ELO anymore .

doubling of time, not doubling of cores. no perfect scaling

I believe the machine has 128 cores. Whether using 250 threads is better than 128 threads on such a machine is questionable, probably the elo gain or lost is miniscule. So seven doublings of cores. Maybe about like four doublings of speed, i would guess.

[Nay Lin Tun]

In CCC one thread SF (NNUE) beats 250 threads Ethereal! And plays equal to 250 threads Komodo or SF classic. Definitely doubling of cores don't give 70 ELO anymore .

It looks like you did not know before that engine strength tapers over time like 'rat tail'

[AndrewGrant]

Any idea what version of Ethereal it was? Guessing it was not an NNUE Ethereal? +120 elo or so.

[Jouni]

Archive says: Ethereal tcec s20 divp (popcnt).

[MikeB]

In CCC one thread SF (NNUE) beats 250 threads Ethereal! And plays equal to 250 threads Komodo or SF classic. Definitely doubling of cores don't give 70 ELO anymore .

An interesting comment ...

Code: Select all

        Source: CCRL 40/15
	Stockfish 13 64-bit 4CPU	3552	+20	−20	77.4%	−183.3	45.2%	892    NNUE	 
	Stockfish 13 64-bit 1CPU        3520	+23	−22	74.5%	−168.2	49.2%	666    NNUE
	Elo delta                         32

 	Stockfish 12 64-bit 4CPU	3516	+19	−19	71.8%	−140.4	56.0%	841    NNUE
	Stockfish 12 64-bit 1CPU        3478	+15	−15	74.5%	−164.2	48.9%	1612   NNUE
	Elo delta                         38

	Stockfish 11 64-bit 4CPU	3478	+19	−19	73.0%	−142.8	52.6%	840    Classical
	Stockfish 11 64-bit 1CPU   	3434	+15	−15	74.0%	−156.3	49.8%	1493   Classical
	Elo delta                         44
	
	Stockfish 10 64-bit 4CPU	3462	+17	−16	73.4%	−155.6	50.3%	1227   Classical
	Stockfish 10 64-bit 1CPU	3386	+9	−9	74.6%	−168.0	46.7%	4812   Classical
	Elo delta                         78
	
	Stockfish 9 64-bit 4CPU	        3430	+15	−15	72.3%	−153.4	50.5%	1423  Classical
	Stockfish 9 64-bit 1CPU	        3365	+12	−12	71.4%	−149.3	49.0%	2565  Classical
	Elo delta                         75
	
	Stockfish 8 64-bit 4CPU     	3377	+14	−13	65.1%	−96.8	62.1%	1647  Classical
	Stockfish 8 64-bit 1CPU     	3300	+11	−11	69.9%	−131.1	53.8%	2670  Classical
	Elo delta                         77
	

From Stockfish 8 to Stockfish 10 - roughly ~75 Elo spread between 4 PU to 1 core
There was big decline going from Sf 10 to SF11, and since then, the NNUE versions show further decline in the spread between 4CPU and 1CPU.

I certainly do not have any answers , but the sharp decline in the spread between 4CPU to 1CPU with going from SF10 to SF11 is very suspicious.
Seems to be more than simply "diminishing returns" ,which we would all expect at some point, buy why all of a sudden such a sharp drop?

Perhaps we are at now at a point now where Elo gains will be very slow going from this point forward? If we are, that is not a good thing for computer chess - as far as human play goes, we have a long way to go yet ;>), a few centuries at least. It may mark the end of Correspondence Chess (Computer Assisted, ala ICCF) as we know it today.

[Cornfed]

Perhaps we are at now at a point now where Elo gains will be very slow going from this point forward? If we are, that is not a good thing for computer chess - as far as human play goes, we have a long way to go yet ;>), a few centuries at least. It may mark the end of Correspondence Chess (Computer Assisted, ala ICCF) as we know it today.

The part in red...can you explain?
Or are you trying to say that those humans with the stronger helpers and more $$ to throw at them...will be no better off really than someone with 1 core?