Last Update: 28.01.2022
Hello Chess Friends,
I am pleased to announce that,I realized to run new Multi-Core eng testing mainly as NNUEAlso just for comparing, I tested SF 310720 NON-NNUE too
The main target of the current Multi-Core testing is that,To test and see the latest Stockfish dev. via Evalfile, plusTo check the influence of 1 Core, 2 Cores, 4 Cores, 8 CoresNote that current/latest matches are played as Ponder OFFPonder ON is much more effective with slower TC: Details
Shocking news! as NNUE no much Elo difference...On other hand, this is not as a BIG surprise...!)And here are the latest NNUE Multi-Core standings:
# PLAYER : RATING POINTS GAMES (%)1 8 Core : 3746.8 448.0 840 53.3%2 4 Core : 3740.2 437.5 840 52.1%3 2 Core : 3728.9 419.5 840 49.9%4 1 Core : 3700.8 375.0 840 44.6%
More Details:• Stockfish 250122 dev (via nn-ac07bd334b62.nnue) is used to play• Parallel matches, but for each player is installed other copy engine • NNUE's overall draw percentage (based on 1680 games) high: 93%• The Database has no any game lost on time, due to played.1m+1s
GAMES
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 1 Core vs 2 Core: The Elo diff: 21Btw, please don't forget to compare the results as NON-NNUE too
1 2 Core +20/-3/=257 53.04% 148.5/2802 1 Core +3/-20/=257 46.96% 131.5/280
• NNUE 1c vs 2c draw perc. (based on 280 games): 92%
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 2 Core vs 4 Core: The Elo diff: 11
1 4 Core +11/-2/=267 51.61% 144.5/2802 2 Core +2/-11/=267 48.39% 135.5/280
• NNUE 2c vs 4c draw perc. (based on 280 games): 95%
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 4 Core vs 8 Core: The Elo diff: 7
1 8 Core +6/-0/=274 51.07% 143.0/2802 4 Core +0/-6/=274 48.93% 137.0/280
• NNUE 4c vs 8c draw perc. (based on 280 games): 98%--------------------------------------------------------------------------------------------
What I can say more over current NNUE engine testings,Actually this is already known issue that since NNUE appeared,Latest SF based ones each other produce close to 99% drawsI mean in case of MP, Ponder ON, Blitz, Strong openings etc.Even under current conditions, we noticed about 93% draws...Be aware of that too, the engines played via 4-MEN Syzygy..If 6-7-MEN..then the current draw percentage would be higherIn other words, not sure how many time will take, but via NNUE:Computer Chess games are expecting to be ended all as drawSure, I am referring for eng matches each other, which are 3700+Btw, who knows?)Perhaps Chess is already solved as MP, Ponder ON, 40/120 ?Meanwhile (as another proof since NNUE appeared on scene)Hardware speeds, opening books don't play BIG role as before..E.g for anyone missed...for more information: Playchess 2021Btw, the users of low-end or mid-end PCs are probably happy..)But unfortunately, it seems the high-end machines are affecting...That's why I say, NNUE revolution is close to Lenin revolution !)As you may see too, via NNUE: the speeds are more equalized!--------------------------------------------------------------------------------------------
And here are the new results played as NON-NNUE
1st NON-NNUE Test: 1 Core vs 2 Core = The Elo diff: 86
1 2 Core +80/-12/=188 62.14% 174.0/2802 1 Core +12/-80/=188 37.86% 106.0/280
More Details:• Stockfish 310720 C40 is used, the same version played: Results• Parallel matches, but for each player installed other copy engine• The current overall draw percentage (based on 280 games): 67%
--------------------------------------------------------------------------------------------
2nd NON-NNUE Test: 2 Core vs 4 Core = The Elo difference: 80
1 4 Core +78/-15/=187 61.25% 171.5/2802 2 Core +15/-78/=187 38.75% 108.5/280
More Details:• Stockfish 310720 C40 is used, I mean the same SF vers is played• As usual, parallel matches, but for each player installed other copy...• Overall draw perc. (based on 280 games): 67% (same as above..)
--------------------------------------------------------------------------------------------
3rd NON-NNUE Test: 4 Core vs 8 Core = The Elo difference: 55
1 8 Core +56/-12/=212 57.86% 162.0/2802 4 Core +12/-56/=212 42.14% 118.0/280
More Details:• Stockfish 310720 C40 is used, I mean the same SF vers is played• As usual, parallel matches, but for each player installed other copy...• The current overall draw percentage (based on 280 games): 76% --------------------------------------------------------------------------------------------
As Final Conclusion between both testings: NNUE Vs NON-NNUE:- Draw percentage difference is really very BIG, approx. 25%- NNUE 1 Core vs 2 Core, Elo difference is minus plus 20- NNUE 2 Core vs 4 Core, Elo difference is minus plus 10- NNUE 4 Core vs 8 Core, Elo difference is minus plus 7- NON-NNUE 1 Core vs 2 Core Elo diff.is minus plus 85 - NON-NNUE 2 Core vs 4 Core Elo dif. is minus plus 80- NON-NNUE 4 Core vs 8 Core Elo dif. is minus plus 55
What I can point out too, as you may know: SF via Evalfile plays stronger, but no much gain from the speeds! Same words are going e.g in case of TC such as 2m+1s vs 8m+1s Btw, in this Test: time control's Elo diff. is close to 1 Core vs 2 Core
And as I stated before (via NNUE about 3700+Elo each other...):, To see a win it's something like looking for a needle in a haystack!) This is one of the reasons (I mean for mostly of SCCT eng testings), About why I lost interest and prefer mainly as NON-NNUE: Details Plus via huge assistance, any Fool can manage to play stronger..!)
By the way, without to not mention this I can not too, I see some Testers don't allow 'Engine Learning' as enabled... But in same time, they are testing the engines via Evalfiles Not same, but these testings lead to close method as another! For this reason I 'm a little bit confused, but no BIG problem )
And to be more clear, 1st of all, and once more, the EvalFiles are not chess engines! In short, NN files are learned data, where NNUE engine gains... Otherwise, I am afraid that we will mix the apples with oranges) Hope helps...--------------------------------------------------------------------------------------------
Btw just I'd like to add the below tests may be useful tooNote also that I ran these engine strength testings for CS XLVII
More Details:The above testing is played as Ponder ON, via nn-cb26f10b1fd9.nnue 3 Cores used to play via 512 MB hash, where 1 Core via 128 MB hash• NNUE 1core vs 3core draw percentage (based on 360 games): 91%*The Elo Difference is 29----------------------------------------------------------------------------------------------Meanwhile, as NON-NNUE: the Elo difference is minus plus 140 Note: The below results are played by Brainfish 151019: DeepBrain
# PLAYER : RATING POINTS GAMES (%) 4 Ponder ON 3CPU : 3639.1 302.0 600 50.3%7 Ponder ON 1CPU : 3500.0 295.5 700 42.2%------------------------------------------------------------------------------------------------
And as a last note,I hope all the current data will be helpful... sure we need more of them...But at least now we have new NON-NNUE vs NNUE stats to compare!
Best,Sedat
Hello Chess Friends,
I am pleased to announce that,I realized to run new Multi-Core eng testing mainly as NNUEAlso just for comparing, I tested SF 310720 NON-NNUE too
The main target of the current Multi-Core testing is that,To test and see the latest Stockfish dev. via Evalfile, plusTo check the influence of 1 Core, 2 Cores, 4 Cores, 8 CoresNote that current/latest matches are played as Ponder OFFPonder ON is much more effective with slower TC: Details
Shocking news! as NNUE no much Elo difference...On other hand, this is not as a BIG surprise...!)And here are the latest NNUE Multi-Core standings:
# PLAYER : RATING POINTS GAMES (%)1 8 Core : 3746.8 448.0 840 53.3%2 4 Core : 3740.2 437.5 840 52.1%3 2 Core : 3728.9 419.5 840 49.9%4 1 Core : 3700.8 375.0 840 44.6%
More Details:• Stockfish 250122 dev (via nn-ac07bd334b62.nnue) is used to play• Parallel matches, but for each player is installed other copy engine • NNUE's overall draw percentage (based on 1680 games) high: 93%• The Database has no any game lost on time, due to played.1m+1s
GAMES
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 1 Core vs 2 Core: The Elo diff: 21Btw, please don't forget to compare the results as NON-NNUE too
1 2 Core +20/-3/=257 53.04% 148.5/2802 1 Core +3/-20/=257 46.96% 131.5/280
• NNUE 1c vs 2c draw perc. (based on 280 games): 92%
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 2 Core vs 4 Core: The Elo diff: 11
1 4 Core +11/-2/=267 51.61% 144.5/2802 2 Core +2/-11/=267 48.39% 135.5/280
• NNUE 2c vs 4c draw perc. (based on 280 games): 95%
--------------------------------------------------------------------------------------------
NNUE Individual Stats, 4 Core vs 8 Core: The Elo diff: 7
1 8 Core +6/-0/=274 51.07% 143.0/2802 4 Core +0/-6/=274 48.93% 137.0/280
• NNUE 4c vs 8c draw perc. (based on 280 games): 98%--------------------------------------------------------------------------------------------
What I can say more over current NNUE engine testings,Actually this is already known issue that since NNUE appeared,Latest SF based ones each other produce close to 99% drawsI mean in case of MP, Ponder ON, Blitz, Strong openings etc.Even under current conditions, we noticed about 93% draws...Be aware of that too, the engines played via 4-MEN Syzygy..If 6-7-MEN..then the current draw percentage would be higherIn other words, not sure how many time will take, but via NNUE:Computer Chess games are expecting to be ended all as drawSure, I am referring for eng matches each other, which are 3700+Btw, who knows?)Perhaps Chess is already solved as MP, Ponder ON, 40/120 ?Meanwhile (as another proof since NNUE appeared on scene)Hardware speeds, opening books don't play BIG role as before..E.g for anyone missed...for more information: Playchess 2021Btw, the users of low-end or mid-end PCs are probably happy..)But unfortunately, it seems the high-end machines are affecting...That's why I say, NNUE revolution is close to Lenin revolution !)As you may see too, via NNUE: the speeds are more equalized!--------------------------------------------------------------------------------------------
And here are the new results played as NON-NNUE
1st NON-NNUE Test: 1 Core vs 2 Core = The Elo diff: 86
1 2 Core +80/-12/=188 62.14% 174.0/2802 1 Core +12/-80/=188 37.86% 106.0/280
More Details:• Stockfish 310720 C40 is used, the same version played: Results• Parallel matches, but for each player installed other copy engine• The current overall draw percentage (based on 280 games): 67%
--------------------------------------------------------------------------------------------
2nd NON-NNUE Test: 2 Core vs 4 Core = The Elo difference: 80
1 4 Core +78/-15/=187 61.25% 171.5/2802 2 Core +15/-78/=187 38.75% 108.5/280
More Details:• Stockfish 310720 C40 is used, I mean the same SF vers is played• As usual, parallel matches, but for each player installed other copy...• Overall draw perc. (based on 280 games): 67% (same as above..)
--------------------------------------------------------------------------------------------
3rd NON-NNUE Test: 4 Core vs 8 Core = The Elo difference: 55
1 8 Core +56/-12/=212 57.86% 162.0/2802 4 Core +12/-56/=212 42.14% 118.0/280
More Details:• Stockfish 310720 C40 is used, I mean the same SF vers is played• As usual, parallel matches, but for each player installed other copy...• The current overall draw percentage (based on 280 games): 76% --------------------------------------------------------------------------------------------
As Final Conclusion between both testings: NNUE Vs NON-NNUE:- Draw percentage difference is really very BIG, approx. 25%- NNUE 1 Core vs 2 Core, Elo difference is minus plus 20- NNUE 2 Core vs 4 Core, Elo difference is minus plus 10- NNUE 4 Core vs 8 Core, Elo difference is minus plus 7- NON-NNUE 1 Core vs 2 Core Elo diff.is minus plus 85 - NON-NNUE 2 Core vs 4 Core Elo dif. is minus plus 80- NON-NNUE 4 Core vs 8 Core Elo dif. is minus plus 55
What I can point out too, as you may know: SF via Evalfile plays stronger, but no much gain from the speeds! Same words are going e.g in case of TC such as 2m+1s vs 8m+1s Btw, in this Test: time control's Elo diff. is close to 1 Core vs 2 Core
And as I stated before (via NNUE about 3700+Elo each other...):, To see a win it's something like looking for a needle in a haystack!) This is one of the reasons (I mean for mostly of SCCT eng testings), About why I lost interest and prefer mainly as NON-NNUE: Details Plus via huge assistance, any Fool can manage to play stronger..!)
By the way, without to not mention this I can not too, I see some Testers don't allow 'Engine Learning' as enabled... But in same time, they are testing the engines via Evalfiles Not same, but these testings lead to close method as another! For this reason I 'm a little bit confused, but no BIG problem )
And to be more clear, 1st of all, and once more, the EvalFiles are not chess engines! In short, NN files are learned data, where NNUE engine gains... Otherwise, I am afraid that we will mix the apples with oranges) Hope helps...--------------------------------------------------------------------------------------------
Btw just I'd like to add the below tests may be useful tooNote also that I ran these engine strength testings for CS XLVII
NNUE test, to see the influence: 1 Core vs 3 Cores:
1 CF EXT 161120 3 Core +31/-1/=328 54.17% 195.0/3602 CF EXT 161120 1 Core +1/-31/=328 45.83% 165.0/360More Details:The above testing is played as Ponder ON, via nn-cb26f10b1fd9.nnue 3 Cores used to play via 512 MB hash, where 1 Core via 128 MB hash• NNUE 1core vs 3core draw percentage (based on 360 games): 91%*The Elo Difference is 29----------------------------------------------------------------------------------------------Meanwhile, as NON-NNUE: the Elo difference is minus plus 140 Note: The below results are played by Brainfish 151019: DeepBrain
# PLAYER : RATING POINTS GAMES (%) 4 Ponder ON 3CPU : 3639.1 302.0 600 50.3%7 Ponder ON 1CPU : 3500.0 295.5 700 42.2%------------------------------------------------------------------------------------------------
And as a last note,I hope all the current data will be helpful... sure we need more of them...But at least now we have new NON-NNUE vs NNUE stats to compare!
Best,Sedat