Texel 1.07

[Jouni]

I tested how Texel plays endings with 5 piece tablebases (from EG_Msb.epd).

Code: Select all

endgame  2017

1	Stockfish 220917 64 BMI2		171.5/300
2	Houdini 6 x64-pext	  		 167.5/300
3	Komodo 10 64-bit	 	    	154.5/300
4	Texel 1.07	 		          106.5/300

So around 100 ELO weaker than top engines.

[petero2]

TC: 30m+5s
Texel107-Cluster [T60] - Texel107-12C: 41.5 - 28.5 (70: +13, =57, -0)
PGN

TC: 10m+3s
Texel107-Cluster [T60] - Texel107-12C: 111 - 69 (180: +43, =136, -1)
PGN

Of course too few games to make clear statements, but it seems that Texel scales well on 60 cores compared to 12 cores.

Thanks for the test results.

Using the standard logistic elo formula the results correspond to +65 elo and +83 elo respectively. Dividing by log2(5) to get the approximate elo increase per doubling of number of cluster nodes gives +28 elo and +36 elo.

This is roughly in line with what I have gotten in my tests at shorter time control.

I have also recently run a test comparing 32 dual core cluster nodes against 16 dual core cluster nodes. (m4.xlarge EC2 instances). My result was +25 elo after 3776 games.

I don't know if these numbers are good or bad since I don't have anything to compare with. The elo increase is a lot less than when doubling time and when doubling the number of cores on a single computer, but that is certainly expected.

I am currently testing Cluster Toga on my Ethernet "ad-hoc" cluster, but I think Cluster Toga may not be that efficient unless you have a low latency cluster using for example InfiniBand.

[reflectionofpower]

Is there some kind of software that will make all 800 parameters visible in Fritz GUI ?

I see 40+ at best in Fritz GUI. No way in H-E-Double hockey sticks will it show 800 BUT you could edit the UCI file IF you knew all the parameters and put them in.

[main line]

What is strongest:
1. Johny on 2000 cores
2. Texel on 2000 cores
3. Houdini 6 on Xeon 64 cores
time control 40/120 min + 20/60min + 45min
If Texel is 120 Elo below top 3 programs in PC with same cores, maybe Texel can be better than Houdini, Komodo or Sf, used by cluster of huge cores for ex. Texel 128 cores against Houdini 12 cores?

[petero2]

What is strongest:
1. Johny on 2000 cores
2. Texel on 2000 cores
3. Houdini 6 on Xeon 64 cores
time control 40/120 min + 20/60min + 45min
If Texel is 120 Elo below top 3 programs in PC with same cores, maybe Texel can be better than Houdini, Komodo or Sf, used by cluster of huge cores for ex. Texel 128 cores against Houdini 12 cores?

I don't know how either Jonny or Texel scales on a large cluster. Texel is around 250 elo below Houdini 6 though on CCRL 40/40. The 100 elo reported earlier in this thread was in special endgame positions when TBs were in use.

It would really surprise me if Texel on 128 cores would be as strong as Houdini on 12 cores.

[Milos]

What is strongest:
1. Johny on 2000 cores
2. Texel on 2000 cores
3. Houdini 6 on Xeon 64 cores
time control 40/120 min + 20/60min + 45min
If Texel is 120 Elo below top 3 programs in PC with same cores, maybe Texel can be better than Houdini, Komodo or Sf, used by cluster of huge cores for ex. Texel 128 cores against Houdini 12 cores?

At that TC doubling of cores on a cluster can hardly be worth more than 25Elo. Going from 12 to 128 cores is 3.2 doublings or around 80Elo at best.
Going from 64 to 2000 cores is 5 doublings but with so many cores that is probably worth less than 25Elo per doubling, i.e. most probably less than 120Elo in total. So no, neither Johny nor Texel are stronger than Houdini with 30 times more cores.

[Dann Corbit]

What is strongest:
1. Johny on 2000 cores
2. Texel on 2000 cores
3. Houdini 6 on Xeon 64 cores
time control 40/120 min + 20/60min + 45min
If Texel is 120 Elo below top 3 programs in PC with same cores, maybe Texel can be better than Houdini, Komodo or Sf, used by cluster of huge cores for ex. Texel 128 cores against Houdini 12 cores?

At that TC doubling of cores on a cluster can hardly be worth more than 25Elo. Going from 12 to 128 cores is 3.2 doublings or around 80Elo at best.
Going from 64 to 2000 cores is 5 doublings but with so many cores that is probably worth less than 25Elo per doubling, i.e. most probably less than 120Elo in total. So no, neither Johny nor Texel are stronger than Houdini with 30 times more cores.

I think you are probably right about that, but it would be fun to see such a contest anyway.

[main line]

But we saw this here:
https://icga.leidenuniv.nl/?page_id=1990
All engines are equal

In 5th round 2 years ago Jonny-2400 cores won Komodo-12 cores
https://icga.leidenuniv.nl/?page_id=1260

I am sure in one thing:
Texel with 2400 cores in such time control ~1hour45min/game + increment 15sec/move is unbeatable because with huge number of cores and in long time control engine will find strongest ("God's") move in any position.
No difference between 2000 cores and 10.000 cores or 100.000 cores because for strongest move in any position with nowadays strogest engines need probably 2400 cores 2-5 minutes / move.

[Dann Corbit]

But we saw this here:
https://icga.leidenuniv.nl/?page_id=1990
All engines are equal

In 5th round 2 years ago Jonny-2400 cores won Komodo-12 cores
https://icga.leidenuniv.nl/?page_id=1260

I am sure in one thing:
Texel with 2400 cores in such time control ~1hour45min/game + increment 15sec/move is unbeatable because with huge number of cores and in long time control engine will find strongest ("God's") move in any position.
No difference between 2000 cores and 10.000 cores or 100.000 cores because for strongest move in any position with nowadays strogest engines need probably 2400 cores 2-5 minutes / move.

I do not think anything significant can be discovered by a tiny contest like Leiden.

I think that Texel's excellent scale power might be worth more than the estimates, but there is no way to tell unless you play the games.

For extreme scaling, Ipman's site is really interesting.
This refers to a slightly older version of Texel (which was a big improvement from the previous version):
http://www.ipmanchess.yolasite.com/test ... hreads.php

Texel 1.06a48 512cores Numatest 3 from Peter Osterlund

He was so kind to make a new release Texel 1.06a49 with source and can put the link here:

https://dl.dropboxusercontent.com/u/896 ... l106a49.7z


setoption name hash value 1024
setoption name threads value 72
go depth 22

8cores
info nodes 523751942 nps 8781890 time 59640

16cores
info nodes 465912687 nps 16656990 time 27971

18cores=1cpu
info nodes 700218378 nps 18361567 time 38135

22cores
info nodes 474150640 nps 20808858 time 22786

32cores
info nodes 915012019 nps 31049985 time 29469

36cores =2cpu's
info nodes 789579580 nps 35225499 time 22415

44cores
info nodes 1054602231 nps 42282183 time 24942

54cores=3cpu's
info nodes 960015462 nps 48781273 time 19680

64cores
info nodes 1113647063 nps 56077700 time 19859

72cores
info nodes 987339559 nps 62092922 time 15901

I did also used go depth 24 for last two ,because it was to quickly finished with go depth 22

64cores - go depth 24
info nodes 4634818942 nps 63032176 time 73531

72cores - go depth 24
info nodes 6291455339 nps 69499644 time 90525


Compare it to this version and yow will see it is about 5x better on 72 cores that the previous:

Texel had a first try with Numa..but also here no gain after 2cpu's..even a slowdown..also SMP can be better!

Texel 1.06a48 512cores Numa

setoption name threads value 8
go depth 22

8cores
info nodes 535900101 nps 7208965 time 74338

16cores
info nodes 666496072 nps 13589203 time 49046

18cores
info nodes 467481214 nps 15624894 time 29919

22cores
info nodes 719367492 nps 17628099 time 40808

32cores
info nodes 906991000 nps 21920702 time 41376

36cores
info nodes 924095974 nps 22428969 time 41201

44cores
info nodes 958960416 nps 15240947 time 62920

54cores
info nodes 1357458581 nps 12538295 time 108265

64cores
info nodes 1145138765 nps 14235759 time 80441

72cores
info nodes 1056645969 nps 14224983 time 74281

[Uri Blass]

But we saw this here:
https://icga.leidenuniv.nl/?page_id=1990
All engines are equal

In 5th round 2 years ago Jonny-2400 cores won Komodo-12 cores
https://icga.leidenuniv.nl/?page_id=1260

I am sure in one thing:
Texel with 2400 cores in such time control ~1hour45min/game + increment 15sec/move is unbeatable because with huge number of cores and in long time control engine will find strongest ("God's") move in any position.
No difference between 2000 cores and 10.000 cores or 100.000 cores because for strongest move in any position with nowadays strogest engines need probably 2400 cores 2-5 minutes / move.

I see no reason to believe that programs will find the strongest move in any position in the near future even if you give them 24 hours per move but note that you do not need it in order to be unbeatable because you need only to find moves that are good enough not to lose in practical game.

Note that I believe that texel with 2400 cores at the time control that you mention is not unbeatable.