Uri Blass wrote: ↑Sun Feb 23, 2025 10:12 am
Considering the fact that stockfish17 is probably better than stockfish15 you may need the hardware to be 100 times slower than 486 to get super GM strength.
Yes. During the pandemic I did a crude experiment (viewtopic.php?t=76820&start=20) with Stockfish 12 and 13 in an emulated Macintosh Quadra (Motorola 68k) playing against very strong dedicated chesscomputers with a somewhat accepted Elo.
During the tests I limited fishs' nodes to come down to 300-500 (Nodes/second) during middlegame and it still did beat the opposition convincingly.
Now (17) with the ultra-strong implementation of NNUE_CYBER_NET, it just will need a fraction of the cpu-power. If you get it running on a toaster/router (smart fridges and television sets are already overpowered...), it will still certainly crush a human grandmaster.
Uri Blass wrote: ↑Sun Feb 23, 2025 10:12 am
Considering the fact that stockfish17 is probably better than stockfish15 you may need the hardware to be 100 times slower than 486 to get super GM strength.
Yes. During the pandemic I did a crude experiment (viewtopic.php?t=76820&start=20) with Stockfish 12 and 13 in an emulated Macintosh Quadra (Motorola 68k) playing against very strong dedicated chesscomputers with a somewhat accepted Elo.
During the tests I limited fishs' nodes to come down to 300-500 (Nodes/second) during middlegame and it still did beat the opposition convincingly.
Now (17) with the ultra-strong implementation of NNUE_CYBER_NET, it just will need a fraction of the cpu-power. If you get it running on a toaster/router (smart fridges and television sets are already overpowered...), it will still certainly crush a human grandmaster.
Have a nice sunday!
I have not previously heard of NNUE_CYBER_NET - what is that and is it something that the public can download to use with Stockfish 17?
royb wrote: ↑Sun Feb 23, 2025 3:20 pm
I have not previously heard of NNUE_CYBER_NET - what is that and is it something that the public can download to use with Stockfish 17?
Thanks!
Hello.
Not to worry - you're using it already.
My assumption would be that the speed of the search of stockfish 17 slowed down to ≈ 250 nodes per second during the middlegame (in the endgame it's higher of course) should be suffice now to reach 2700 "Elo".
royb wrote: ↑Sun Feb 23, 2025 3:20 pm
I have not previously heard of NNUE_CYBER_NET - what is that and is it something that the public can download to use with Stockfish 17?
Thanks!
Hello.
Not to worry - you're using it already.
My assumption would be that the speed of the search of stockfish 17 slowed down to ≈ 250 nodes per second during the middlegame (in the endgame it's higher of course) should be suffice now to reach 2700 "Elo".
greetings
By 2700 "Elo", do you mean "should score 50% in classical Time limit standard chess against humans currently rated about 2700 FIDE"? I have no opinion about that as of now, but I would like to know what evidence there is for such a human Elo rating in classical chess for SF? Is it perhaps based on playing against ancient software that was known to be about that level twenty years ago, such as Fritz, Junior, Rebel, or Shredder versions? If so I would point out that 2700 FIDE today is about like 2800 FIDE was then, based on Accuracy studies. Any evidence of the human elo of a specific node speed on SF 17 would be very useful information.
204975 <Stockfish 12 m68k@45MHz(0): info depth 14 seldepth 19 multipv 1 score cp 64 nodes 69599 nps 394 hashfull 62 tbhits 0 time 176276 pv d2d4 d7d5 c2c4 e7e6 b1c3 g8f6 c1g5 f8e7 e2e3 f6e4 g5e7 c6e7 f1d3 e4c3 b2c3
204975 <Stockfish 12 m68k@45MHz(0): bestmove d2d4 ponder d7d5
204975 >Tasc R30 The King 2.5(1): position startpos moves g1f3 b8c6 d2d4
204975 >Tasc R30 The King 2.5(1): isready
So depth 13/14 was enough for the wiping. As I said "crude experiment"...
Any evidence of the human elo of a specific node speed on SF 17 would be very useful information.
Yes. Node-buckets don't work reliably I think - that's why I opted for heavy slow-down.
If I'd be inclined to experiment again, I'd repeat my setup but with the several maja-builds (instead of the dedicated comps) this time.
204975 <Stockfish 12 m68k@45MHz(0): info depth 14 seldepth 19 multipv 1 score cp 64 nodes 69599 nps 394 hashfull 62 tbhits 0 time 176276 pv d2d4 d7d5 c2c4 e7e6 b1c3 g8f6 c1g5 f8e7 e2e3 f6e4 g5e7 c6e7 f1d3 e4c3 b2c3
204975 <Stockfish 12 m68k@45MHz(0): bestmove d2d4 ponder d7d5
204975 >Tasc R30 The King 2.5(1): position startpos moves g1f3 b8c6 d2d4
204975 >Tasc R30 The King 2.5(1): isready
So depth 13/14 was enough for the wiping. As I said "crude experiment"...
Any evidence of the human elo of a specific node speed on SF 17 would be very useful information.
Yes. Node-buckets don't work reliably I think - that's why I opted for heavy slow-down.
If I'd be inclined to experiment again, I'd repeat my setup but with the several maja-builds (instead of the dedicated comps) this time.
A 99% score is about a +800 elo result with the logistic formula, so the rating could be as high as 3150 based on that test! Clearly for a fairly accurate rating (as opposed to a minimum rating), it needs to run against something with an SSDF rating of around 2700 (I think the SSDF list is accurate enough to estimate human classical elo equivalent). How many nodes per move or nodes per second on SF 17 can break even with an SSDF 2700 at 90' + 30"? That's actually not too hard to test.
How many nodes per move [...] on SF 17 can break even with an SSDF 2700 at 90' + 30"?
Wouldn't want to know.
How many [...] nodes per second on SF 17 can break even with an SSDF 2700 at 90' + 30"?
Maybe slightly more interesting (Nodes must not be arbitrarily fixed!!!!!111111!!!!1; slow down the speed and let stockfish figure out the nodes during different phases of the game; one has to agree to/begin at a base) - but then it would be just for the case of Stockfish (a moving target) 17. Gets old rather soon.
That's actually not too hard to test.
No it isn't. You could easily do it with a computer.