Not logged inRybka Chess Community Forum
Forum RybkaChess Help Search Register Login
Up Topic Rybka Support & Discussion / Rybka Discussion / Questions for Vas regarding upcoming Deep Rybka 5
1 2 3 Previous Next  
- - By Magnus Friedmann (***) [ro] Date 2012-01-09 15:07
Happy New Year Vas and welcome back! :) Do you plan to sell Rybka 5 under Chessbase with a new powerful book? It would be awesome! How many cores has the latest Rybka Cluster? Thank you. I'm looking forward for Deep Rybka 5 to be released, and I strongly believe it will regain the number 1 spot as it has always deserved! All the best!
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-09 15:14

> Do you plan to sell Rybka 5 under Chessbase with a new powerful book?


Yes and yes (but it's not a concrete plan yet).

> How many cores has the latest Rybka Cluster?


Lukas can give you the exact amount. It's around 300.

Vas
Parent - - By Magnus Friedmann (***) [ro] Date 2012-01-09 15:28
Thank you very much Vas for the provided information! All the best!
Parent - - By Permanent Brain (*****) Date 2012-01-09 15:40
Which information?:grin:
Parent - By Dragon Mist (****) [hr] Date 2012-01-09 16:45
:lol:
Parent - - By Werewolf (*****) [gb] Date 2012-01-09 17:13
Let me try to ask some more testing questions: :smile:

i) Will Rybka 5 be geared for multi-socket (NUMA) machines like Houdini 2 Pro?
ii) Will you release a single core version for IDeA use?
iii) In Rybka's development a couple of years ago you stated that you spent roughly 2/3 of your time on search and 1/3 on evaluation. Has that proportion changed or stayed the same?
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-09 21:37
i) No, I doubt it.
ii) Yes (100% sure).
iii) Maybe it's more like 3/4 search and 1/4 eval now. It's hard to say for sure, a lot of my work is not specific to either.

Vas
Parent - By Master Om (*****) [in] Date 2012-05-13 16:30
Will you add King Safety Eval Parameter ?
Will it contain Learning feature like critter ?
Parent - - By Banned for Life (Gold) Date 2012-01-09 19:58
A hot question here on the forum is how well your cluster algorithms work with a large number of cores. Care to speculate on what kind of improvement in Elo you would get from doubling the cores from ~300 to ~600?
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-09 21:39
We'd probably get a 1.6x speedup going from 300 to 600 cores. What that translates to in Elo terms will depend on test conditions.

Vas
Parent - - By Banned for Life (Gold) Date 2012-01-09 22:15
Wow. A 1.6X real speadup going from 300 to 600 cores without shared memory is in the 'unbelievable' realm. Not that I would bet against you in a freestyle game! :lol:
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-09 22:21
Please note that that's a WAG. It might easily be 1.5x. But it's decent.

Vas
Parent - - By Labyrinth (****) [us] Date 2012-01-10 00:16
If you got an average of 100 kN/sec per core and could run on a supercomputer with 100,000 cores, you'd achieve a paltry 10 gN/sec.

Anyways I have heard that Rybka's kN/sec output was "deliberately obfuscated". I thought the reasoning was that you could really only compare two identical engines to get an accurate benchmark anyways, so Rybka's node count was made the way it was so it would be more convenient for development. Clarification?
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-10 08:36
Rybka's node counts are lower than how other engines count nodes (AFAIK).

Rybka's node counts are also adjusted to take into account for multiprocessing inefficiency. If you have two Rybka processes and they each search 100 nodes, Rybka will report around 170 nodes searched rather than 200, because the value of the combined work is about the same as a single 170-node search. I don't know how other engines handle this.

Vas
Parent - - By Banned for Life (Gold) Date 2012-01-10 08:55
Other engines count every node without accounting for MP inefficiency. This has led some to conclude (incorrectly) that Rybka scales worse than some other engines...
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-10 11:51
Aha. Maybe the NPS scaling should be superlinear with # of processors. :smile:

Vas
Parent - - By Banned for Life (Gold) Date 2012-01-10 15:16
With H2, NPS increases almost exactly linearly out to 12 (16?) nodes...
Parent - - By Werewolf (*****) [gb] Date 2012-01-10 15:20

> With H2, NPS increases almost exactly linearly out to 12 (16?) nodes...


So then you need to allow for a 20% loss of efficiency per doubling of cores.

http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=23899

Seems like Vas has built this 20% into his node counter.
Parent - - By Banned for Life (Gold) Date 2012-01-10 15:23
Vas has built a derating curve into his reported NPS values, but I hope it is based on something more substantial than just a 20% loss of efficiency per doubling...
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-10 17:06
Actually that's pretty much the formula. I think it's 1.7x speedup per doubling.

Vas
Parent - - By Stonehenge (***) Date 2012-01-10 20:54
Indeed, 20% efficiency reduction corresponds to 2 / 1.2 = 1.67x effective speed per doubling, pretty much the same value.
Apparently your hardcore fans are quite disappointed that it's as simple as that. :lol:
Parent - - By NATIONAL12 (Gold) [gb] Date 2012-01-11 02:10
i knew it was approx 1.7 nobody is disappointed in this,did you think otherwise.Trust you were amused by my 40/40 testing K4-H2

http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=23839;pg=1
Parent - - By Arrière Pensée (Gold) Date 2012-01-11 02:42 Edited 2012-01-11 02:44
You didn't access Gaviota Tb's ? Did you? I didn't see any Tb hits on any of the games accessed by Houdini. That's why I'm asking. Or, maybe that wouldn't effect the outcome of any of the games played.
Parent - - By NATIONAL12 (Gold) [gb] Date 2012-01-11 02:49
no.
Parent - By Arrière Pensée (Gold) Date 2012-01-11 02:53
The only reason I'm asking is because I don't have a clue how to run a single core match and I did note that on IPON they utilitize 64 MB hash and 4 piece tbs.
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-11 07:33

> Apparently your hardcore fans are quite disappointed that it's as simple as that. :lol:


I guess Alan expected me to account for some other things like interprocess latencies.

Vas
Parent - - By Banned for Life (Gold) Date 2012-01-13 07:47
I thought you (or Lukas) actually did some representative measures of speedup going from 1 to 2 to 4 to 8 to 12 cores and tried to fit some curve to these values. I'm not sure I believe the speedup in going from 2 to 3 cores is the same as going from 8 to 12. I certainly don't believe that the speedup from going from 1 to 2 cores would be the same as going from 32 to 64 cores.
Parent - By Vasik Rajlich (Silver) [pl] Date 2012-01-13 08:01
Aha. My formula is of course an approximation. There are a lot of things which genuinely influence the speedup which it doesn't consider. Still, it is surprisingly accurate.

Vas
Parent - - By suj (***) Date 2012-01-10 15:48
Some do account for MP/cluster inefficiency.
Parent - - By Banned for Life (Gold) Date 2012-01-10 15:49
Which ones?
Parent - - By suj (***) Date 2012-01-10 15:51
Sjeng (Cluster) does. You can actually see the effiency and measure it.
Parent - By Banned for Life (Gold) Date 2012-01-10 15:56
OK, I was kind of referring to available engines. :wink:
Parent - By Arrière Pensée (Gold) Date 2012-01-09 22:33
Oh, my God! You need to get out more!
Parent - - By Arrière Pensée (Gold) Date 2012-01-10 06:59
Of course my only request will be - 

Will you commit to a bugfix update! Do we get the "showstopper" option thrown in again? Just for old times sake. :smile:
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-10 08:37
Actually I've decided to stop having big bugs in my code. :smile:

It was a close call, but ...

Vas
Parent - - By Arrière Pensée (Gold) Date 2012-01-10 20:37
Aah! The influence of experiencing an expanding universe and the changing of diapers?:grin:
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-10 23:24
More the former than the latter. :smile:

Vas
Parent - - By Arrière Pensée (Gold) Date 2012-01-10 23:53
I've noted that the Crown is that of the Queen and not of the King -was that intentional?
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-11 07:37
I don't know. Maybe Dadi wanted to use a knight and accidentally uploaded the wrong file! :smile:

I see you got your gold crown after all. :surprised:

Vas
Parent - - By Arrière Pensée (Gold) Date 2012-01-11 07:51
My point being the next in line could be the king's crown !:razz::smile: But I suppose that might be too confusing at this point.
Parent - - By TheHug (Bronze) [us] Date 2012-01-11 08:15
I guess Uly is the only one in line to get that one right now. :smile:
Parent - - By Arrière Pensée (Gold) Date 2012-01-11 08:31
Actually, maybe he would be a good one to come up with a design for the next symbol.
Parent - - By TheHug (Bronze) [us] Date 2012-01-11 08:32
(I did read your last post)

I think all the ideas are good to be honest. Uly is great at design ideas. :smile:
Parent - - By Uly (Gold) [mx] Date 2012-01-11 17:00
I always wanted a knight in the ranks, but now that there's a crown already there...

Oh, and I never got the bronze or silver crowns because when they were implemented I jumped to gold directly. I was thinking if I could have a second forum account (with permission) to go for them.
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-11 17:13
Maybe the crowns should be displayed side by side. For example, if you have enough posts for a gold crown plus enough posts left over for another bronze crown, both should be displayed.

Vas
Parent - - By Uly (Gold) [mx] Date 2012-01-11 18:22
Well, that makers me think about the solution for further ranks:

25000 posts - 2 Gold crowns
50000 posts - 3 Gold crowns
75000 posts - 4 Gold crowns
100000 posts - 5 Gold crowns

:lol:
Parent - - By Vasik Rajlich (Silver) [pl] Date 2012-01-11 18:39
That sounds pretty steep. Isn't one gold crown 10,000 posts? In this case I would say that 20,000 should be 2, 30,000 posts should be 3, etc.

Vas
Parent - - By Uly (Gold) [mx] Date 2012-01-11 18:56
Hmm, the steepness could be fixed with intermediate ranks:

20,000 Gold + Bronze
25,000 Gold + Silver
30,000 Gold + Gold
35,000 Gold + Gold + Bronze
40,000 Gold + Gold + Bronze
45,000 Gold + Gold + Silver
50,000 Gold + Gold + Gold
55,000 Gold + Gold + Gold + Bronze
60,000 Gold + Gold + Gold + Silver
65,000 Gold + Gold + Gold + Gold
70,000 Gold + Gold + Gold + Gold + Bronze
75,000 Gold + Gold + Gold + Gold + Silver
80,000 Gold + Gold + Gold + Gold + Gold

Though it doesn't end on the mighty 100k.

(I hope this isn't KISS :lol:)
Parent - - By turbojuice1122 (Gold) [us] Date 2012-01-11 21:46
Oh, you just want more crowns for yourself. :razz:
Parent - By Uly (Gold) [mx] Date 2012-01-11 21:48
Well, I'm thinking 20 years down the road :wink:
Up Topic Rybka Support & Discussion / Rybka Discussion / Questions for Vas regarding upcoming Deep Rybka 5
1 2 3 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill