TalkChess.com

mclane wrote: ↑Sat Oct 05, 2019 11:28 pm

If you have a car that has 100 PS and drives good, it does not make much sense to drive another car that has 200 Ps or 300 Ps.

It does if you are in a car race. It does not make sense if you do not believe in car races but just want a comfortable jouney.
You are like a person driving at 30 mph shouting at cars in a racing rally. What is the point of this speed. I am more comfortable than you.

wrote: Instead of programming the 10-20 elo points you could buy a new computer that is twice as fast, that would give the same progress.

Well people want more and more elos for the same computer. Programmers supply what the market wants. You think there should not be such a market as the market is pointless.

wrote: If you don’t change the method, nothing will change at all.

Current method is least energy spent for most elo. Why change the method unless you are a purist masochist.

Did you ever read one of Asimovs robot books ?

Wow, I just read this entire thread for the first time. It seems someone decided to celebrate the thread's 6 month decease by turning it into a zombie.

It seems the claim is that as hardware improvements exploded, coders forgot about chess, and instead of figuring out how to make engines play better chess, they focused in making use of those hardware improvements in the most efficient way.

The claim makes logical sense:

Imagine a world where hardware never got to the multi-core stage, then it follows that all chess programmers that spent time implementing MP in their engines instead would have used that time improving the engine in other ways. In a world where 64bit computers never existed, instead of programming for 64bit, programmers would have used their time making 32bit programs stronger, etc.

Suppose then that we could communicate with those worlds, and arranged some interchange of source code. Since they don't have multicore machines or 64bit, we'd have to send them the strongest thing we have that runs on their machine, and they'd laugh their asses by how pitiful it performs compared to what they have. On the other hand, it'd be trivially easy to get their source code, rewrite the concepts so that it runs on 64bit, and add the best multicore code that we can, and since their engine would understand chess better, it'd outperform whatever we had today.

The fact would be that if hardware improvements halted to a point where Stockfish could only analyze 1000 nodes per move on a regular basis, then computer chess would have taken a very different course, and in that world, improvements would have needed to come from somewhere else, and who knows if we'd have had a breakthrough where engines would play as strong as today's engines by just looking at a fraction of the nodes.

Stockfish limited to 1000 nodes per move:
[pgn] [Event "Computer Schach Partie"] [Site "ORION8"] [Date "2019.04.05"] [Round "?"] [White "Mephisto III S Glasgow"] [Black "Stockfish_10_x64_popcnt"] [Result "1:0"] [ECO "B22"] [Opening "Sizilianisch (Alapin 2.c3), 2...Nf6 3.e5 Nd5 4.d4 cxd4 5.Nf3 Nc6 6.Bc4"] [Time "21:53:52"] [Variation "(Alapin 2.c3), 2...Nf6 3.e5 Nd5 4.d4"] [WhiteElo "2400"] [Termination "normal"] [PlyCount "50"] [WhiteType "human"] [BlackType "program"] 1. e4 c5 2. c3 Nf6 3. e5 Nd5 4. d4 cxd4 5. Qxd4 Nc7 {-1.63/7 0} 6. Bc4 b5 {+0.78/7 0} 7. Bb3 e6 {-0.69/7 0} 8. Nf3 Nc6 {-1.46/6 0} 9. Qe4 h6 {-0.17/6 0} 10. O-O a5 {-0.82/6 0} 11. a4 b4 {-1.43/6 0} 12. Rd1 bxc3 {+0.31/7 0} 13. Nxc3 Rb8 {-1.11/6 0} 14. Bc4 Ba6 {-1.95/6 0} 15. Bxa6 Nxa6 {-1.79/7 0} 16. Qe2 Nc5 {-0.12/7 0} 17. Be3 Nb3 {-0.94/7 0} 18. Rab1 Rb4 {-1.25/7 0} 19. Qc2 Be7 {-0.57/7 0} 20. Nb5 O-O {+0.54/6 0} 21. Rd3 Rxa4 {-3.05/6 0} 22. Rxb3 Nb4 {-2.87/6 0} 23. Qe4 Qb8 {-4.78/6 0} 24. Nc3 d5 {-9.62/6 0} 25. exd6 Qxd6 {-9.38/7 0 Schwarz gibt auf} 1:0 [/pgn]

Mephisto III S Glasgow does ~ 3-5 NPS and runs on a 12 MHz 68000. Maybe due to wait states it is only effectively ~ 8 MHz.

Just a small match with 1 node per move for Leela, versus Stockfish dev with 3000 nodes per move, with 100 predefined varying ECO positions:

George Tsavdaris wrote: ↑Mon Oct 07, 2019 1:25 pm Just a small match with 1 node per move for Leela, versus Stockfish dev with 3000 nodes per move, with 100 predefined varying ECO positions:

Code: Select all

1   Leela lc0v22.0,T40B.4-160, 1 node  +133/-43/=24 72.50%  145.0/200
2   Stockfish 190923 3000 nodes        +43/-133/=24 27.50%   55.0/200

This is very interesting. I have always hoped for a rating list that showcased such entities and their relative strengths.

Like, do you know who would win in a match of Komodo Depth 20 vs. Stockfish Depth 25? Does anybody have any idea about it or a way to estimate the result? No, because nothing like this has been tested, so a good start would be some CCRL equivalent engine at 40/40 against Stockfish at some fixed depth, and then that fixed depth against nodes per move, and then nodes per move against Leela at 1 node. And then Komodo at fixed depth could play any of those and we'd have an idea of what's what.

At least, this sounds a lot more interesting than what engine testers are doing.