Vas Speaks in Tongues

Code, algorithms, languages, construction...
BB+
Posts: 1484
Joined: Thu Jun 10, 2010 4:26 am

Re: Vas Speaks in Tongues

Post by BB+ » Wed Jan 25, 2012 11:49 am

... he already made lots of money very quickly from the long hard work of others that will likely never be realized in computer chess again.
I think this is exactly one of FL's complaints. For some reason the Rybka Forum likes to imagine that somehow the ICGA (or maybe the cloners) have "ruined" VR, forced him out of the business, or whatever -- when they overlook the actual fact that this is exactly what happened when Rybka appeared as a "Fruit derivative".

I don't have all the details, but the famed 051103 version of Fruit (dated Nov 3 2005 but didn't appear until years later -- and looks just as strong as R1, if not better) might have been held up in its release due to the licensing contract talks with Convekta, and when these were scuppered by Rybka appearing, by that time FL had more pressing personal issues to which to attend, while Benitez took a different development path. It is only speculation, but perhaps with an inital cash infusion for serious testing equipment, Fruit would have gained something similar to the 280 Elo (exaggerated to 400 by various) between R1 and R4.1 over the next years.

Put another way, the use of Fruit-derived code in R1 already provided a huge boost in the world of opportunities to VR (ignoring any complaints about later versions), and FL was never able to get his foot in the door again.
Chris, Ed and RF posters have refuted exactly nothing. Just a great waste of computer network bandwidth and nothing else...
I do always get the feeling the RF is an alternate reality (though it is chuckly that the Houdini Opening Book is advertised from ChessOK on the front page). Such as the recent thread that started: The TalkChess forum has always permitted public viewing of all discussions on that forum. Until today. Proceeding in that thread I find:
Mr. Watkins trashy rebuttal starts by pretending that Rybka violates Fruit's copyright.
Must be a different rebuttal that the one I wrote?! Nowhere do I make such a comment, at the start or anywhere else. I mention copyright law for a couple of paragraphs in the first section (and a few more times later) because it formed a framework for various of the Panel discussions. The commenter seems not to address any of many other issues discussed in my "trashy rebuttal".

I was hoping for some substantive discussion of the points raised by the Riis series and its rebuttal(s), but I gave up trying to find such on the RF. I did find some more kooky claims:
And I have the impression Mark W. also is withholding us better annotated ASM code.
I think the error is that of the writer, who seems to assume that my ASM annotation (of the Rybka eval) stood by itself; rather, it was meant to be read in conjunction with (say) the more verbose EVAL_COMP descriptions, more of as a check that said descriptions were indeed correct.

ernest
Posts: 247
Joined: Thu Sep 02, 2010 10:33 am

Re: Vas Speaks in Tongues

Post by ernest » Wed Jan 25, 2012 6:48 pm

BB+ wrote:the famed 051103 version of Fruit ...looks just as strong as R1, if not better
Are you sure about that?
See CCRL 40/40: Fruit 2.3.1 is rated stronger, and R1 64-bit even stronger: +49 Elo
and in CCRL 40/4: R1 64-bit is +64 Elo to Fruit 051103.

But true, Fruit 051103 was only 32-bit, and not so far away from R1.

BB+
Posts: 1484
Joined: Thu Jun 10, 2010 4:26 am

Re: Vas Speaks in Tongues

Post by BB+ » Wed Jan 25, 2012 7:21 pm

This was my previous post. http://www.open-chess.org/viewtopic.php ... =60#p10902
I used CCRL 40/4 there, but as you say, the margin is wider in 40/40.
2885 Rybka 1.0 32-bit
2870 Fruit 2.1
And here is CEGT 40/4
367	Fruit 051103        	 2827
380	Rybka 1.0 Beta w32 1CPU	 2820
I use 32-bit for Rybka, as that was the main market for which Fruit was being developed (as noted by VR in 2008?, the single-cpu 32-bit version was outpacing the other versions in sales by about 4:1). I think saying they are "equal" is quite reasonable, and if one counts the NIL-42 ratings, Fruit is somewhat better.

ernest
Posts: 247
Joined: Thu Sep 02, 2010 10:33 am

Re: Vas Speaks in Tongues

Post by ernest » Wed Jan 25, 2012 9:23 pm

BB+ wrote:I use 32-bit for Rybka,
OK, but then let's be fair: for some reason, Vasik never cared to do a good job with the 32-bit versions.
The 64/32 speed ratio of Vasik's first versions is x1.8 (still x1.7 for Rybka 4), corresponding to more than +50 Elo.

Only Cozzie's Zappa had a 64/32 speed ratio even higher than x1.8
For today's Houdini and Stockfish, the ratio is more like x1.3

hyatt
Posts: 1242
Joined: Thu Jun 10, 2010 2:13 am
Real Name: Bob Hyatt (Robert M. Hyatt)
Location: University of Alabama at Birmingham
Contact:

Re: Vas Speaks in Tongues

Post by hyatt » Thu Jan 26, 2012 1:15 am

ernest wrote:
BB+ wrote:I use 32-bit for Rybka,
OK, but then let's be fair: for some reason, Vasik never cared to do a good job with the 32-bit versions.
The 64/32 speed ratio of Vasik's first versions is x1.8 (still x1.7 for Rybka 4), corresponding to more than +50 Elo.

Only Cozzie's Zappa had a 64/32 speed ratio even higher than x1.8
For today's Houdini and Stockfish, the ratio is more like x1.3

I have never personally seen such high numbers (1.8x and such).

Unless the code is written in such a way that it needs a ton of registers, the speed difference ought to be down around 1.2x to 1.3x. If the extra 8 x86_64 registers help enough to get it to 1.8x, something is BADLY wrong in that program...

Crafty is a classically optimized bitboard program, and 1.3x was about the right speedup even when comparing 32 bit vs 64 bit on other platforms like the dec alpha...

BB+
Posts: 1484
Joined: Thu Jun 10, 2010 4:26 am

Re: Vas Speaks in Tongues

Post by BB+ » Thu Jan 26, 2012 9:41 am

The 64/32 speed ratio of Vasik's first versions is x1.8
The first data point I have for Rybka 64-bit speedup is "around 20%" (though possibly 40% upon fixing some bug?), see http://www.stmintz.com/ccc/index.php?id=356880
I agree that later versions enlarge this ratio.

ernest
Posts: 247
Joined: Thu Sep 02, 2010 10:33 am

Re: Vas Speaks in Tongues

Post by ernest » Fri Jan 27, 2012 7:47 pm

hyatt wrote:I have never personally seen such high numbers (1.8x and such).
I am a bit surprised at your "ivory tower" type of knowledge.
Even if you never wanted to get dirty by having a look at Rybka, Zappa, ...and Windows, you must have had open-minded students who discussed the subject with you!

If you look at my post below, you will see (for the first time :o ) your x1.8 to x1.9

ernest
Posts: 247
Joined: Thu Sep 02, 2010 10:33 am

Re: Vas Speaks in Tongues

Post by ernest » Fri Jan 27, 2012 8:10 pm

BB+ wrote:The first data point I have for Rybka 64-bit speedup is "around 20%"
Hi BB+,

I just calculated the speed ratio again (in console mode, default hash=32 MB):

Rybka v1.0 Beta.x64.exe 04/12/2005 00:26
Rybka v1.0 Beta.w32.exe 04/12/2005 00:31

Windows XP Pro x64

uci
id name Rybka 1.0 Beta 32-bit
id author Vasik Rajlich
.................
uciok
go depth 15
.................
info depth 15 score cp 7 time 55892 nodes 10365952 nps 191962 pv b1c3 g8f6 g1f3
d7d5 d2d4 g7g6 c1f4 f8g7 g2g3 f6h5
.................
info depth 15 time 68063 nodes 12618284 nps 191186
info time 68063 nodes 12618284 nps 191186
bestmove b1c3 ponder g8f6


uci
id name Rybka 1.0 Beta ( ==> 64-bit)
id author Vasik Rajlich
.....................
uciok
go depth 15
.....................
info depth 15 score cp 7 time 29486 nodes 10365952 nps 370212 pv b1c3 g8f6 g1f3
d7d5 d2d4 g7g6 c1f4 f8g7 g2g3 f6h5
.....................
info depth 15 time 35892 nodes 12618284 nps 360522
info time 35892 nodes 12618284 nps 360522
bestmove b1c3 ponder g8f6

==> speed ratio: x1.90


Windows XP Home

uci
id name Rybka 1.0 Beta 32-bit
id author Vasik Rajlich
………………
uciok
go depth 15
…………………….
info depth 15 score cp 7 time 55439 nodes 10365952 nps 191962 pv b1c3 g8f6 g1f3
d7d5 d2d4 g7g6 c1f4 f8g7 g2g3 f6h5
…………………….
info depth 15 time 67533 nodes 12618284 nps 194127
info time 67533 nodes 12618284 nps 194127
bestmove b1c3 ponder g8f6

==> speed ratio: x1.88

MoldyJacket
Posts: 77
Joined: Sun Mar 20, 2011 1:11 am
Location: Charlotte, NC
Contact:

Re: Vas Speaks in Tongues

Post by MoldyJacket » Fri Jan 27, 2012 10:15 pm

What hardware are you running these tests on?
Nominal Quoting Fanboy

hyatt
Posts: 1242
Joined: Thu Jun 10, 2010 2:13 am
Real Name: Bob Hyatt (Robert M. Hyatt)
Location: University of Alabama at Birmingham
Contact:

Re: Vas Speaks in Tongues

Post by hyatt » Fri Jan 27, 2012 11:30 pm

ernest wrote:
hyatt wrote:I have never personally seen such high numbers (1.8x and such).
I am a bit surprised at your "ivory tower" type of knowledge.
Even if you never wanted to get dirty by having a look at Rybka, Zappa, ...and Windows, you must have had open-minded students who discussed the subject with you!

If you look at my post below, you will see (for the first time :o ) your x1.8 to x1.9

What is "ivory tower knowledge"?

What were the circumstances? Nothing running for both? Are you running the 32 bit code on the same box as the 64 bit code? There are issues doing so, compatibility libraries, wrapper library functions, etc...

The usual test is a 32 bit compiler, a 32 bit O/S, and a 32 bit compiled executable. Ditto for 64 bit stuff. 1.8x-1.9x is VERY high. Much of a bitboard program is NOT 64 bits. The only places we see an advantage with 64 bits is either (a) manipulating the bitboards themselves since we now actually work with 64 bit values using just one instruction rather than a pair of 32 bit instructions; (b) 8 extra registers in 64 bit mode allows the compiler to optimize a bit better.

I'm not so sure that mixing 32 and 64 bit in the way you tested is showing just the hardware gain. You might be showing some software "loss" by running a 32 bit app in a 64 bit environment...

Too many variables to be able to interpret your data. My tests were run on 32 bit and then on 64 bit environments, separately, using the same version of the compiler but one for 32 and one for 64 bits...

I can probably make a 32 bit crafty on my current hardware and see what that does...

But in any case, your numbers seem too high to be solely the result of 64 bit instructions alone...

Post Reply