As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
Stockfish scaling - which cpu equals carlsen level
Moderators: hgm, Rebel, chrisw
-
- Posts: 2821
- Joined: Fri Sep 25, 2015 9:38 pm
- Location: Sortland, Norway
-
- Posts: 12540
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: Stockfish scaling - which cpu equals carlsen level
I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
- Posts: 291
- Joined: Wed May 08, 2013 6:49 am
Re: Stockfish scaling - which cpu equals carlsen level
Stockfish 1,000 CPUs with some hash bugs = Carlsen.
-
- Posts: 13447
- Joined: Wed Mar 08, 2006 9:02 pm
- Location: Dallas, Texas
- Full name: Matthew Hull
Re: Stockfish scaling - which cpu equals carlsen level
I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
It might be interesting to find out what human Elo is required to match its play.
Matthew Hull
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: Stockfish scaling - which cpu equals carlsen level
I estimate 2600 GM.mhull wrote:I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
It might be interesting to find out what human Elo is required to match its play.
-
- Posts: 1010
- Joined: Thu Sep 01, 2011 2:49 pm
Re: Stockfish scaling - which cpu equals carlsen level
Connect this setup to FICS, EICS, ICC or other chess site and let it play against rated humans there. Would be interesting to see its rough rating, as most chess sites these days have calibrated their online chess rating.I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.
It might be interesting to find out what human Elo is required to match its play.
Depth 12 at 3 minutes would be between 2100 to 2300 I guess.
-
- Posts: 13447
- Joined: Wed Mar 08, 2006 9:02 pm
- Location: Dallas, Texas
- Full name: Matthew Hull
Re: Stockfish scaling - which cpu equals carlsen level
Here are the actual numbers from a test game against Crafty 24.1. Crafty was set at depth=11, on my laptop, with book. Cfish ran on m68k (bogomips=4.55, 32-bit, @20Mhz, hash=2Mb) no book.Laskos wrote:I estimate 2600 GM.mhull wrote:I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
It might be interesting to find out what human Elo is required to match its play.
Cfish Performance
Average Depth = 13.1628
Average NPS = 446.946
Average Time = 198.498 (3.3083 minutes)
Game (Annotations by Crafty):
[pgn][Event "Computer Chess Game"]
[Site "HULLMATT1"]
[Date "2017.04.12"]
[Round "-"]
[White "Cfish m68k @20Mhz"]
[Black "Crafty-24.1 depth=11"]
[Result "1-0"]
[TimeControl "40/7200"]
[Annotator "7... +0.74"]
1. e4 c6 2. d4 d5 3. e5 Bf5 4. Nf3 e6 5. Be2 Nd7 6. O-O Ne7 7. Be3 Qb6
{+0.74/11 0.1} 8. Qc1 f6 {+0.77/11 0.2} 9. exf6 gxf6 {+0.76/11 0.1} 10. Nh4
O-O-O {+0.31/11 0.2} 11. c4 Rg8 {+0.33/11 0.2} 12. Nc3 Bh3 {+0.09/11 0.3}
13. Kh1 Nf5 {+0.15/11 0.1} 14. gxh3 Nxh4 {+0.08/11 0.1} 15. cxd5 exd5
{+0.91/11} 16. Nxd5 Qa5 {+0.84/11} 17. Bc4 Rg7 {+0.70/11} 18. Bd2 Qa4
{+1.49/11 0.1} 19. Bh6 Rg5 {+1.94/11} 20. Bxg5 fxg5 {+1.92/11} 21. a3 Nc5
{+2.38/11 0.2} 22. Nb6+ axb6 {+2.21/11} 23. dxc5 Bxc5 {+2.41/11} 24. b3 Qa8
{+2.72/11} 25. Qxg5 Ng6 {+2.90/11} 26. h4 Kc7 {+2.99/11 0.4} 27. f4 Bd4
{+2.66/11 0.2} 28. f5 Ne5 {+2.69/11 0.2} 29. Rad1 Nxc4 {+3.45/11 0.2} 30.
bxc4 c5 {+3.29/11} 31. f6 b5 {+4.12/11} 32. f7 b6+ {+3.83/11} 33. Qg2 Qxg2+
{+4.02/11} 34. Kxg2 Rf8 {+4.18/11} 35. cxb5 Kd6 {+4.20/11 0.1} 36. Kg3 h5
{+4.29/11 0.1} 37. Kg2 Kd5 {+4.55/11 0.1} 38. Rde1 Bg7 {+5.08/11} 39. Rf5+
Kd6 {+4.65/11} 40. Re4 Bh8 {+5.70/11 0.1} 41. Re8 Bg7 {+6.28/11} 42. Rxf8
Bxf8 {+5.33/11 0.1} 43. Rf6+ Ke7 {+6.30/11}[/pgn]
This build of Cfish was a bit wobbly with respect to time management. Using Level 40 120 0, it played too quickly in the opening (10 to 20 seconds/move) and got smashed. Same with with level 20 60 0. The game here was 1 3 0, and it played a bit too slow (3.3 minutes/move), breaking the time control. Time control was set in polyglot on the command line using the Level command.
Matthew Hull
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Stockfish scaling - which cpu equals carlsen level
do arrange playing a single LTC (2-3 hours for 40 moves) game for Carlsen against SF, and the human will achieve twice better results in comparison to those that would ensue if one game is played every single day.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
computers do not get tired, you know.
-
- Posts: 4889
- Joined: Thu Mar 09, 2006 6:34 am
- Location: Pen Argyl, Pennsylvania
Re: Stockfish scaling - which cpu equals carlsen level
Keeping the old Mac SE busy eh? My guess is that it would be 1900-2000 FIDE or add about 100 or so for USCF - with only 400 nodes , a human can see the horizon better than SF in some positions , not all and sometimes it's just one bad move that can kill a game - things change a lot by the time when you get to 5000 nodes searches - at that level Sf would play at at least ~2200 /2300 or so 9- just my opinion, not saying it's factmhull wrote:I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
It might be interesting to find out what human Elo is required to match its play.
-
- Posts: 13447
- Joined: Wed Mar 08, 2006 9:02 pm
- Location: Dallas, Texas
- Full name: Matthew Hull
Re: Stockfish scaling - which cpu equals carlsen level
Mac IIsi 68030 @20Mhz. (A Mac SE has a 68000 chip running @8Mhz).MikeB wrote:Keeping the old Mac SE busy eh?mhull wrote:I'm running Cfish at about 400 NPS (on a 20Mhz, 32-bit machine), average depth, 12 at 3 minutes/move.Dann Corbit wrote:I think it is one of those "How many angels can dance on the head of a pin?" things.Nordlandia wrote:As per title.
Assuming sf9 dev running on 100 Kn/s is apprixmately equal to 2830 level?
The Elo of chess engines comes from a different pool (there is very little cross pollination between the groups) so the real measure of a human player verses a computer player has a huge error bar.
I think that the human strengths and the computer strengths are different.
I guess that high end hardware will beat every GM handily in a long contest. But I am not even sure what that means.
I don't think that the GMs are keen to play thousands of rated games against computers in a public setting so I guess that we will never really know what the strength differences are.
But we will debate it anyway.
It's kind of fun, after all.
It might be interesting to find out what human Elo is required to match its play.
Cfish on Mac IIsi, level 1 3 0 performs as follows:
Average Depth = 13.1628
Average NPS = 446.946
Average Time = 198.498 (3.3083 minutes)
That's interesting because Kai estimates 2600 Elo at 3 minutes per move, average depth 13, at 447 NPS. I cannot yet network this machine. But if I do, I'll try to get it running on a chess server to play actual people.MikeB wrote:My guess is that it would be 1900-2000 FIDE or add about 100 or so for USCF - with only 400 nodes , a human can see the horizon better than SF in some positions , not all and sometimes it's just one bad move that can kill a game - things change a lot by the time when you get to 5000 nodes searches - at that level Sf would play at at least ~2200 /2300 or so 9- just my opinion, not saying it's fact
Or if someone has a computer account on FICS they'd let me operate manually to collect some slow games with humans, we have some real data points to analyse.
Matthew Hull