back


28097Nobody
2018-08-21 03:10:39
@queen13
Scaling test for SF (low number of games, large error bars) :
http://tests.stockfishchess.org/tests/view/5b391fae0ebc5902b2e60162
384 vs 192 threads (both 192 cores) : +22 Elo (so HT is a gain)

http://tests.stockfishchess.org/tests/view/5b3906d70ebc5902b2e5fec5
384 vs 64 threads (both 192 cores, equivalent to 64 cores for the 64 threads limit) : +94 Elo

http://tests.stockfishchess.org/tests/view/5ad998020ebc5902a3efed74
384 vs 32 threads (both 192 cores, equivalent to 32 cores for the 32 threads limit) : +114 Elo

As you can see, even with large error bars, scaling is excellent up to (at least) 384 threads.

About Leela strength : given the way GPUs are evolving (btw thanks @Mrert for the news), Leela will probably getting stronger faster than SF (not even counting better nets for the NN), but as of now, SF is still the king of chess engine (or the queen if you prefer).
28096Mrert
2018-08-20 20:40:19
New Nvidia GPUs unveiled (20X0 line)
28095queen13
2018-08-20 18:01:38
@Nobody
Very interesting.

I thought stockfish could support up to 128 threads

but it used 384 in the test?

two doublings from 48 cores to 192 cores is 90 elo?

really? I thought stockfish(and chess engines in general not just stockfish)

don't scale that well with that many cores.

Did they find a way ? that's awesome !

I thought leela was stronger than stockfish in extreme conditions like these

that there are extreme conditions after which leela dominates

According to you this is not the case and I believe you.

Nothing better than a test of course.
28094Mrert
2018-08-20 17:30:13
@queen13: in theory, the scaling is optimal. But there are always hardware limitations if you have several cpus/gpus working in parallel.
28093queen13
2018-08-20 15:07:10
@Mrert

I thought it scaled perfectly up to a huge number of gpus/cpus...

Now you say they need a good multi gpu code...

or is it just about cpus that it scales perfectly

but for gpus it can support 8-10 and then what?

then it is inefficient to the extent where there is no benefit to

using more or it just doesn't support using more at all?

Thanks a lot for you responses !
28092WCN
2018-08-20 13:28:08
Stockfish 9 (The A.R.B Chess System) Destroys The Latest Leela Chess Zero ID10901 (Hardcore)

https://groups.google.com/forum/#!topic/fishcooking/Wo8sBMi5fNM


https://image.ibb.co/hxZwoK/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_1.jpg

https://image.ibb.co/mQq0he/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_2.jpg

https://image.ibb.co/hK9FFz/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_3.jpg



[Event "LCZero v Stockfish 9 (TARBCS) - 2018.08.19 - Deep Analysis"]
[Date "2018.08.19"]
[Round "1"]
[White "Leela Chess Zero"]
[Black "Stockfish 9 (TARBCS)"]
[Result "0-1"]
[Opening "The A.R.B Chess System"]
[TimeControl "(Hardcore Nodes 2000) 3-5 Seconds Per Move"]
[Termination "normal"]

1. e4 e6 2. Nc3 d6 3. d4 {last book move} h6 4. Bd3 (4. Nf3 Nf6 $14) 4... a6
{Secures b5} (4... e5 5. Nf3 exd4 6. Nxd4 $14) 5. f4 (5. Nf3 Nf6 $14) 5... g6
6. Nf3 b6 (6... Bg7 $5 $14 {might be a viable alternative}) 7. f5 $16 Ne7
(7... exf5 8. Qe2 Ne7 9. O-O $16) 8. O-O (8. f6 Nec6 9. O-O Nd7 $18) 8...
exf5 $16 9. exf5 (9. Qe2 f4 10. Bxf4 Bg7 $18) 9... Bxf5 10. Bxf5 (10. Nh4 Bxd3
11. cxd3 Bg7 $16) 10... Nxf5 11. Qe2+ Ne7 12. Bg5 (12. Qe4 Nbc6 13. d5 $14)
12... Bg7 $11 {Black is behind in development.} (12... hxg5 $2 {looks attractive,
but} 13. Nxg5 Rh5 14. Nce4 $18) 13. Bh4 g5 14. Bxg5 hxg5 15. Nxg5 Bxd4+ 16. Kh1 Be5
17. Nxf7 Nf5 18. Qg4 {White traps the enemy king in the center} (18. Nxh8 $4 {
a gruesome error} Qh4 19. Qxe5+ dxe5 20. Rxf5 Qxh8 $19) 18... Ng3+ 19. Kg1 Qd7
20. Qc4 ({Less advisable is} 20. Nxe5 Qxg4 21. Nxg4 Nxf1 22. Rxf1 Nd7 $17)
20... b5 ({Not} 20... Nxf1 21. Nxh8 d5 22. Qe2 $16) 21. Qd3 Nxf1 {
Pawn h2, live or die!} 22. Nxh8 {Traps the king in the center} Ne3 (22... Nxh2
$142 23. Qg6+ Kf8 24. Qh6+ Qg7 25. Ng6+ Kg8 26. Qxg7+ Bxg7 27. Kxh2 Bxc3 28.
bxc3 Kf7 $15) 23. Re1 $2 (23. Nd5 $142 $5 {would keep White in the game} Kd8
24. Nxe3 (24. Qxe3 $143 Qh7 (24... Bxh8 $143 25. Rf1 Bg7 26. Qg5+ Kc8 27. Ne7+
Kb7 28. Qxg7 $16) 25. Qg5+ Kc8 26. Ne7+ Kb7 $15) 24... Bxh8 25. Qe4 $11) 23...
Nc6 $17 (23... Bxh8 {is the weaker alternative} 24. Qd5 (24. Qxe3+ $2 {
is not as innocent as it looks} Be5 25. Nd5 Qf7 $19) (24. Rxe3+ $6 Be5 25. Qd5
Qc6 26. Qe6+ Kd8 $11) 24... Be5 25. Qxa8 $16 (25. Rxe3 $6 Qc6 26. Qe6+ Kd8 $11)
) 24. Qg6+ Kf8 25. Nd1 Nd5 26. Rf1+ Nf6 27. Nc3 $4
{the position was bad, and this mistake simply hastens the end} (27. Rxf6+ $142
Bxf6 28. Qxf6+ Kg8 29. Ng6 $19) 27... Qg7 $19 28. Nd5 Qxh8 29. Nxf6 (29. Rxf6+
{praying for a miracle} Bxf6 30. Nxf6 $19) 29... Qxh2+ 30. Kf2 Qf4+ 31. Ke1
Qc1+ 32. Kf2 Qf4+ 33. Ke1 Qc1+ 34. Kf2 Qd2+ 35. Kg1 Qd4+ 36. Rf2 Qd1+ 37. Rf1
Bd4+ 38. Kh2 Qxf1 39. Qg8+ Ke7 40. Nd5+ Kd7 41. Qg4+ Ke8 42. Qg8+ Qf8 43. Nxc7+
Kd7 44. Qh7+ Ne7 45. Nxa8 Qf4+ 46. g3 Qf2+ 47. Kh3 Qf1+ 48. Kg4 Qd1+ 49. Kf4
Bxb2 50. Nb6+ Ke8 51. Qe4 (51. Ke3 {does not help much} Qe1+ 52. Kf4 Bc1+ 53.
Kf3 Qf1+ 54. Kg4 Qd1+ 55. Kh3 Qh1+ 56. Kg4 Qxh7 57. Kf3 Qxc2 58. Nc8 Qxc8 59.
a3 Nd5 60. a4 Ne3 61. g4 Qxg4+ 62. Kf2 Qg2+ 63. Ke1 Qd2#) 51... Bc1+ 52. Qe3
Qf1+ 53. Ke4 (53. Kg4 {is not much help} Bxe3 54. Kh4 Ng6+ 55. Kg4 Ne5+ 56. Kh5
Qh1#) 53... Qf5+ 54. Kd4 Qc5+ 55. Kd3 (55. Ke4{cannot undo what has already been done}
Qxe3#) 55... Qxe3# 0-1




WCN
28091Nobody
2018-08-20 12:05:44
Some precisions :
Lc0 with 2x1080 Ti is somewhat at the level of SF 8 *48 cores*
Even with crazy hardware for lc0 *8 V100s*, SFdev *384 threads/192 cores* should still be ahead by like 60 Elo.
28090Nobody
2018-08-20 12:03:39
@queen13 (28088) "then what is stronger? stockfish 128 cores or lc0 8 gpus?"

There is no definitive answer to that but I can give you my guess if you want :
Lc0 with 2x1080 Ti is somewhat at the level of SF 8 (probably a little weaker)
SFdev is 105 Elo stronger than SF8 (approx.)
Latest test (see FishTest) shows that SF with 384 threads/192 cores is 90 Elo stronger than SF 48 cores (which is incredible if you ask me, the scaling is very good for a A/B engine).

Let's be generous and assume that Lc0 can scale with 8 GPUs the same way than SF (x4 power = +90 Elo), which I think is not true, since especially with big NN the memory will become a bottleneck quite early. Let's assume you use V100 instead of 1080 Ti (x2 power = +45 Elo approx.), (8 V100s is worth more than 64'000 $ btw).

That's a lot of approximations and it would obviously need confirmation by actually playing a match between the two.

Even with crazy hardware for lc0, SFdev should still be ahead by like 60 Elo.
28089Mrert
2018-08-19 21:28:41
@queen13: coordinating multiple GPUs are no simple task due to their enormous bandwidth. Gamers complained many years about micro stuttering in SLI/Crossfire setups.
So they need a very good multi GPU code, like SF devs worked a lot on multi CPU code. It's not self evident.
28088queen13
2018-08-19 21:24:07
@Nobody

I thought it could support much more than

"just" 8 great gpus.

if it can support "only" 8 gpus

then what is stronger? stockfish 128 cores or lc0 8 gpus?
28087WCN
2018-08-19 18:56:29
Stockfish 9 (The A.R.B Chess System) Destroys The Latest Leela Chess Zero ID10901 (Hardcore)

https://groups.google.com/forum/#!topic/fishcooking/Wo8sBMi5fNM


https://image.ibb.co/hxZwoK/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_1.jpg

https://image.ibb.co/mQq0he/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_2.jpg

https://image.ibb.co/hK9FFz/LCZero_v_Stockfish_9_TARBCS_2018_08_19_Position_3.jpg



[Event "LCZero v Stockfish 9 (TARBCS) - 2018.08.19"]
[Date "2018.08.19"]
[Round "1"]
[White "Leela Chess Zero"]
[Black "Stockfish 9 (TARBCS)"]
[Result "0-1"]
[Opening "The A.R.B Chess System"]
[TimeControl "(Hardcore Nodes 2000) 3-5 Seconds Per Move"]
[Termination "normal"]

1. e4 e6 2. Nc3 d6 3. d4 h6 4. Bd3 a6 5. f4 g6 6. Nf3 b6 7. f5 Ne7 8. O-O
exf5 9. exf5 Bxf5 10. Bxf5 Nxf5 11. Qe2+ Ne7 12. Bg5 Bg7 13. Bh4 g5 14.
Bxg5 hxg5 15. Nxg5 Bxd4+ 16. Kh1 Be5 17. Nxf7 Nf5 18. Qg4 Ng3+ 19. Kg1 Qd7
20. Qc4 b5 21. Qd3 Nxf1 22. Nxh8 Ne3 23. Re1 Nc6 24. Qg6+ Kf8 25. Nd1 Nd5
26. Rf1+ Nf6 27. Nc3 Qg7 28. Nd5 Qxh8 29. Nxf6 Qxh2+ 30. Kf2 Qf4+ 31. Ke1
Qc1+ 32. Kf2 Qf4+ 33. Ke1 Qc1+ 34. Kf2 Qd2+ 35. Kg1 Qd4+ 36. Rf2 Qd1+ 37.
Rf1 Bd4+ 38. Kh2 Qxf1 39. Qg8+ Ke7 40. Nd5+ Kd7 41. Qg4+ Ke8 42. Qg8+ Qf8
43. Nxc7+ Kd7 44. Qh7+ Ne7 45. Nxa8 Qf4+ 46. g3 Qf2+ 47. Kh3 Qf1+ 48. Kg4
Qd1+ 49. Kf4 Bxb2 50. Nb6+ Ke8 51. Qe4 Bc1+ 52. Qe3 Qf1+ 53. Ke4 Qf5+ 54.
Kd4 Qc5+ 55. Kd3 Qxe3# 0-1





WCN
28086noname
2018-08-19 16:34:27
I can't do testings as my crappy laptop has no GPU so I don't know the answers to your questions queen13 😕, but TCEC published this...

>>
The traditional chess engines will once again run on a 44 cores computer.
The NN engines will be provided with 2x GTX 1080 Ti GPU hardware (on a remote machine).
According to experts in the TCEC chat, this is approximately 35% of the power that Google supplied to Alpha Zero.
<<

Remembering that Alpha Zero played an underpowered SF 8 without appropriate settings.

Inferring then that 6x GTX 1080 Ti GPUs could be enough for an approx 👿equallish👿 match ❓

But I have seen reports that 10x GTX 1080 Ti GPUs have been insufficient to beat SF in a match (with a much earlier and weaker Leela version) ❓

Leela (lc0) seems much stronger now.
28085Nobody
2018-08-19 16:12:36
@queen13 (28079)
Lc0 supports at least up to 8 GPUs.
Apparently in TCEC (2 GPU 1080 Ti) one Gpu was overheating the other, they had to put the fans at 100% and limit the cards to 70% of their maximum to obtain stable games. After that, Lc0 scored 4,5/5 in 5 games, and finished ultimately third of the division.
So Lc0 probably deserved a shot in div 2, for what it's worth.
They played some bonus games after div3 with Lc0 against Ethereal and sf8 (all draw), so maybe Lc0 is somewhat at the level of SF8 (probably a little weaker) *at TCEC's conditions*.
Better luck next season, lc0 could maybe reach divp.
28084WCN
2018-08-19 11:19:54
The reason the so called learning (NN) Chess engines are doomed to Fail...


HOW MANY POSSIBLE MOVE COMBINATIONS ARE THERE IN CHESS?

http://www.bernmedical.com/blog/how-many-possible-move-combinations-are-there-in-chess

There are only 64 spaces and there are 16 pieces each. How many possible move combinations are there?
Chess is infinite: There are 400 different positions after each player makes one move apiece. There are 72,084 positions after two moves apiece. There are 9+ million positions after three moves apiece. There are 288+ billion different possible positions after four moves apiece. There are more 40-move games on Level-1 than the number of electrons in our universe. There are more game-trees of Chess than the number of galaxies (100+ billion), and more openings, defences, gambits, etc. than the number of quarks in our universe! --Chesmayne

Even when we start with something so simple, we can see how quickly complexity explodes.




WCN
28083Albert
2018-08-19 10:38:31
@Mrert, 28082,

I do not think it was because of Deux that Lc0 did not get into Division 2 !!
28082Mrert
2018-08-19 08:10:59
@noname: Yup, hope TCEC will have the guts to ban DeusX. I know there's a hype with NN engines but still plagiarism is unacceptable
28081WCN
2018-08-18 23:30:12
MINDS MASTERING MACHINES [M³] THE ML AND AI CONFERENCE...

https://www.mcubed.london




WCN
28080Albert
2018-08-18 20:11:01
Jon Mike
(another idiot from Lzero Forum) wrote:

Lc0 10800 played excellent games vs SF 8! While she is watching all 6 games, she displayed here superiority.
The games show she is already slightly stronger than SF8! Keep it up Lc0 community! Soon Leela wants to surpass AZ!

I'm laughing. Hahahahahaha ...

All fanatics of Lzero Forun can not stand the Lz0 despite 2x Gtx 1089ti miserably in division 3 failed.

Hahahahahaha ...
28079queen13
2018-08-18 15:50:28
so if I understand correctly...

even at massive hardware like tcec

lc0 is still weaker, much weaker?

someone should find the tipping point !

The point from which on stronger hardware will mean that lc0

is the strongest chess engine.

How many gpus can it currently support?
28078noname
2018-08-18 11:41:22
Immensely relieved that DeusX didn't make it ☺
ASilver won't be getting rich on the backs of other person's hard work (for now) 😀
28077Mrert
2018-08-18 04:55:34
Very interesting ideas from our maintainer Stéphane
28076WCN
2018-08-17 22:54:24
Stockfish 9 (The A.R.B Chess System) Destroys Leela Chess Zero

https://groups.google.com/forum/#!topic/lczero/r5KTdcDW_P8



WCN
28075noname
2018-08-17 22:28:05
They're here

http://tcec.chessdom.com/dl.php?season=13&bo=1
28074Mrert
2018-08-17 22:19:14
I think there are 3 versions:
- CPU (openBLAS ...)
- GPU general (openCL...)
- Nvidia GPU (CUDA, CudNN)
28073WCN
2018-08-17 22:19:02
28065 Albert
2018-08-17 21:36:40 I am very happy that the Lz0 not in the Div. 2 ascended.

After all, she had 2 !! Gtx1080ti available and still it was not enough.

It also got quiet in the Lz0 forum.

No Edward Panek, Jesse Jordache, Chris Withington or Dietrich Kappe, so the fanatics, now speak out and would like to disappear in the ground.

So there is justice in life.


WCN


Very true Albert :) it's lucky the honest people in the world can get a good nights sleep without having to worry about such things as "How can I pull off another TCEC SCAM!" Etc.


WCN

next page >
single page