APassionForCriminalJustic wrote:
Houdini will absolutely have no advantage in this superfinal. There is no need to be concerned with an engine that was 250+ Elo behind Stockfish. It does not matter what opening you throw out there - the advantage for Stockfish will not change. With Komodo out of the way in this Season 9 Stockfish will mop the floor up in what I believe will be the most lopsided superfinal as of yet. Let the games do themselves justice now.
APassionForCriminalJustice-
At the time Houdini 4 was around Stockfish itself was 250 elo lower than it is now, so this logic doesn't hold.
For all we know Houdart could have been quietly working away for the last 2 years making a much stronger version. And the latest Houdini is WAY stronger than Houdini 4.
Lyudmil Tsvetkov wrote:Nice to hear about the 44-core upgrade. With the VLTC this will boost strength somewhat.
But, I prefer to have a 100-elo stronger version and play on single core rather than having a 100-elo weaker engine playing on 44 cores.
I suspect SF 8 single core will win SF 7 16 cores and draw SF 7 44 cores in a 100-game match, or?
You are dreaming. SF 7 16 cores is at least 150Elo stronger than SF 7 single core. SF 8 is current SFdev which is less than 80Elo (most probably around 60Elo) stronger compared to SF7. Further more SF 7 on 44 cores is at least 200Elo stronger than SF 7 single core. So we might see SF10 surpasses SF 7 on 44 cores, but even that is optimistic.
APassionForCriminalJustic wrote:
Houdini will absolutely have no advantage in this superfinal. There is no need to be concerned with an engine that was 250+ Elo behind Stockfish. It does not matter what opening you throw out there - the advantage for Stockfish will not change. With Komodo out of the way in this Season 9 Stockfish will mop the floor up in what I believe will be the most lopsided superfinal as of yet. Let the games do themselves justice now.
APassionForCriminalJustice-
At the time Houdini 4 was around Stockfish itself was 250 elo lower than it is now, so this logic doesn't hold.
For all we know Houdart could have been quietly working away for the last 2 years making a much stronger version. And the latest Houdini is WAY stronger than Houdini 4.
Correct. It always gets me that when you do not see someone putting out a new version of their engine every 6 mos then it has to be that they have dropped the ball on the world of programming.
"Without change, something sleeps inside us, and seldom awakens. The sleeper must awaken." (Dune - 1984)
Lyudmil Tsvetkov wrote:Nice to hear about the 44-core upgrade. With the VLTC this will boost strength somewhat.
But, I prefer to have a 100-elo stronger version and play on single core rather than having a 100-elo weaker engine playing on 44 cores.
I suspect SF 8 single core will win SF 7 16 cores and draw SF 7 44 cores in a 100-game match, or?
You are dreaming. SF 7 16 cores is at least 150Elo stronger than SF 7 single core. SF 8 is current SFdev which is less than 80Elo (most probably around 60Elo) stronger compared to SF7. Further more SF 7 on 44 cores is at least 200Elo stronger than SF 7 single core. So we might see SF10 surpasses SF 7 on 44 cores, but even that is optimistic.
I am not that certain.
That migth have been true 5 or 10 years ago, but not now. I do not think anyone has run similar tests until now. Performance boost based on speed/cores and LTC tends to drop the stronger an engine is, but not so qualitative eval/search/time management changes. It migth be the case the figures you quote have less in common with truth than the figures I am guessing: as said, no one has run similar tests until now.
It would be interesting, when SF 8 appears, for someone to test SF8 - 16 cores vs SF 7 - single core.
[quote="Rochester"]No share the system in the weird way like before.[/quote]
Using of two cluster or two simple PC this is the main benefit.
I think a test with two PC (one PC for one engine and other PC for other engine) produces more reliable result then a primitive but more cheaper test based on one PC. In particular the tests of engines for PC with multiprocessor would require this test method.
It is very pity that while the chess engines have advanced in great degree but in the area of test rules ancient and very confused conditions.
I wonder at engine developers that they take notice of this without any words.
Lyudmil Tsvetkov wrote:That migth have been true 5 or 10 years ago, but not now. I do not think anyone has run similar tests until now. Performance boost based on speed/cores and LTC tends to drop the stronger an engine is, but not so qualitative eval/search/time management changes. It migth be the case the figures you quote have less in common with truth than the figures I am guessing: as said, no one has run similar tests until now.
It would be interesting, when SF 8 appears, for someone to test SF8 - 16 cores vs SF 7 - single core.
So according to you 44 cores is worth less than 80Elo, and just to remind you not on LTC, since this 80Elo gain of SF8 vs SF7 is at bullet TC?
I.e. less than 15Elo per doubling?
Sorry but you don't have a clue what you are talking about. Keep dreaming.
Yes, it all boils down to the opening set, provided the engines are of roughly equal strength, as in current TCEC final.
I can pick 100 positions where SF will score 90% vs Komodo, and another 100, where SF will score only 10% against Komodo. The engine's search and eval parameters will be well tuned for one set, and badly tuned or inexisting for the other.
In a randomly chosen set of 100 opening positions, 60 migth favour Houdini parameters and 40 SF parameters, or vice-versa, but any bigger skew of randomness would be simply bizarre. The split migth also be 50-50, but we do not know that.
That concerns only engines of roughly equal strength. If one engine is some 100 elo better than another, than of course the opening set will have only a minor influence on the score, but that hypothesis is discarded in current TCEC.
Lyudmil Tsvetkov wrote:That migth have been true 5 or 10 years ago, but not now. I do not think anyone has run similar tests until now. Performance boost based on speed/cores and LTC tends to drop the stronger an engine is, but not so qualitative eval/search/time management changes. It migth be the case the figures you quote have less in common with truth than the figures I am guessing: as said, no one has run similar tests until now.
It would be interesting, when SF 8 appears, for someone to test SF8 - 16 cores vs SF 7 - single core.
So according to you 44 cores is worth less than 80Elo, and just to remind you not on LTC, since this 80Elo gain of SF8 vs SF7 is at bullet TC?
I.e. less than 15Elo per doubling?
Sorry but you don't have a clue what you are talking about. Keep dreaming.
So let's wait for SF 8 to be released and test. I do not have a 16-core, but someone certainly would be able to do it.
My claim would be that 100-elo-stronger engine the level of current SF will beat its opponent in a 100 game match, 1 vs 16 cores, so, if SF 8 is only 80 elo stronger, than my claim would be that it will beat SF 7 on 12 cores.
