Rybka 3: The dark truth behind the hype

[Albert Silver]

I just finished the match between Rybka 3 64-bit and Fritz 11, using the Silver Suite. The idea was to test the top two single-CPU contenders against each other and see what came of it. Although on some lists Naum 3.1 is even slightly ahead of Fritz 11, its results against Rybka 3 were pretty bad, so Fritz 11 seemed like the strongest contender under the circumstances. Frankly, the results are hard to quantify. Just so you know, all this prelude is to sort of try and help reduce the shock.

Code: Select all

Rybka 3 		+75/-4/=21	85.5/100
Fritz 11		+4/-75/=21	14.5/100

I won't try to tell you how much better this is over Rybka 232a. Suffice it to say this is a performance of a full +308 Elo over Fritz 11. Seventy-five wins to four... It kind of boggles the mind. If this commentary sounds like hyperbole, let me ask you how you would describe it.

Here is a cute game it played, where a seemingly innocuous queenless middlegame (at least to poor innocent me... ) quickly turned into a mate attack:

1. d4 d5 2. c4 c6 3. Nf3 Nf6 4. Nc3 e6 5. e3 Nbd7 6. Bd3 dxc4 7. Bxc4 b5 8. Bd3 Bb7 9. O-O {end of book}

[D]r2qkb1r/pb1n1ppp/2p1pn2/1p6/3P4/2NBPN2/PP3PPP/R1BQ1RK1 b kq - 0 9

This is one of the positions from the Silver Suite. There are several acceptable continuations. Each engine plays it as both white and black.

9...a6 10. e4 c5 11. d5 c4 12. Bc2 Qc7 13. Nd4 Nc5 14. Qe2 b4 15. Na4 O-O-O 16. Nc6 Bxc6 17. dxc6 Qxc6 18. Qxc4 Qb5

[D]2kr1b1r/5ppp/p3pn2/1qn5/NpQ1P3/8/PPB2PPP/R1B2RK1 w - - 0 19

Although I can find the attacking moves that followed the queen exchange with other engines, none of them chose to exchange queens here as Rybka 3 did.

This kind of brought to mind a comment by the late Rudolph Spielman, a great tactician, who said that he was perfectly capable of finding every single combination Alekhine found, but as to setting up those positions that led to the combinations....

19. Qxb5 axb5 20. Nxc5 Bxc5

[D]2kr3r/5ppp/4pn2/1pb5/1p2P3/8/PPB2PPP/R1B2RK1 w - - 0 21

So queens are now off, and sure, white has an edge, but I didn't expect it to continue as it did

21. a3 b3 22. Bxb3 Nxe4 23. a4 Kb7 24. Bc2 Nf6 25. axb5 Kb6

[D]3r3r/5ppp/1k2pn2/1Pb5/8/8/1PB2PPP/R1B2RK1 w - - 0 26

26. b4! Smack! Bxb4 (if 26...Bd4 27. Ra6+ and if Kxb5 28. Bd3+ is crushing)

[D]3r3r/5ppp/1k2pn2/1P6/1b6/8/2B2PPP/R1B2RK1 w - - 0 27

27. Ra6+! Kxb5 28. Ra7! Kc4 29. Rc7+ Bc5

[D]3r3r/2R2ppp/4pn2/2b5/2k5/8/2B2PPP/2B2RK1 w - - 0 30

and here is a challenging (winning) move for engines that Rybka 3 finds in about 1 second

30. Bd1! Rd7 31. Be2+ Kb4 32. Rc6 Ne4 33. Be3 Ka3 34. Bxc5+ Nxc5 35. Rb6 Nb3 36. Bc4 Kb2 37. Bxb3 1-0

[Event "3m + 2s blitz - Silver Suite"]
[Site "?"]
[Date "2008.07.27"]
[Round "?"]
[White "Rybka 3"]
[Black "Fritz 11"]
[Result "1-0"]
[ECO "D48"]
[PlyCount "73"]
[EventDate "2008.07.27"]
[SourceDate "2008.07.27"]

1. d4 d5 2. c4 c6 3. Nf3 Nf6 4. Nc3 e6 5. e3 Nbd7 6. Bd3 dxc4 7. Bxc4 b5 8. Bd3 Bb7 9. O-O {end of book} a6 10. e4 c5 11. d5 c4 12. Bc2 Qc7 13. Nd4 Nc5 14. Qe2 b4 15. Na4 O-O-O 16. Nc6 Bxc6 17. dxc6 Qxc6 18. Qxc4 Qb5 19. Qxb5 axb5 20. Nxc5 Bxc5 21. a3 b3 22. Bxb3 Nxe4 23. a4 Kb7 24. Bc2 Nf6 25. axb5 Kb6 26. b4 $1 Bxb4 27. Ra6+ $1 Kxb5 28. Ra7 Kc4 29. Rc7+ Bc5 30. Bd1 Rd7 31. Be2+ Kb4 32. Rc6 Ne4 33. Be3 Ka3 34. Bxc5+ Nxc5 35. Rb6 Nb3 36. Bc4 Kb2 37. Bxb3 1-0

[Uri]

Each core of course is 4.8 GHz.
8 x 4.8 GHz = 38.4 GHz! Impressive right?

Is there such a thing in the market yet? Where can i find a 4.8 GHz core?

[M ANSARI]

Each core of course is 4.8 GHz.
8 x 4.8 GHz = 38.4 GHz! Impressive right?

Is there such a thing in the market yet? Where can i find a 4.8 GHz core?

I use a custom made dual vapor cooling system. This allows me to drop the temperature to -40C on each processor ... which allows the processor to clock from 3.2 Ghz to 4.8 Ghz. You would also need to get a motherboard that supports overclocking such as the Intel D5400XS (Skulltrail) and unlocked Extreme Edition processors such as the QX9775. Obviously it is a little more complicated than that since you also have to do things that will prevent condensation and other cooling related problems at such low temperatures.

[M ANSARI]

Thanks Albert for posting these results ... it confirms that R3 is no hype. The results of single core and up seem to show fairly consistent results ... I have seen statistically similar results at 3 0 and 16 0 and even 60 1 time controls ... R3 is a beast and in a few weeks this will be confirmed by all the independent testers. I don't think there has ever been a release of an engine that is so much stronger than the rest. This is especially interesting since I always thought that you couldn't get much stronger than 2.3.2a. By the way how is Brazil these days ... I miss it a lot.

[Uri]

I use a custom made dual vapor cooling system. This allows me to drop the temperature to -40C on each processor ... which allows the processor to clock from 3.2 Ghz to 4.8 Ghz. You would also need to get a motherboard that supports overclocking such as the Intel D5400XS (Skulltrail) and unlocked Extreme Edition processors such as the QX9775. Obviously it is a little more complicated than that since you also have to do things that will prevent condensation and other cooling related problems at such low temperatures.

Cool. Although i don't understand how it works because my knowledge of computer hardware is next to zero. How does dual vapor cooling work?

[bob]

So 1 minute. I'm a bit puzzled by this since i thought an octal Skultraill QX9775 overclocked, as yours, would be around 14-15 times faster than Albert's 7.5 minutes solution that was on a 2.2GHz Athlon64.

An MP program generally has to look at more nodes to evaluate a position than an SP program. This is because cutoff values aren't known precisely. I have read that, in order to make comparing hardware strength easier, Rybka doesn't count the extra nodes the MP program searches.

There is also a lot of randomness in time it takes the MP program to find the critical move, so you can't project the speedup with any accuracy.

Yes when comparing time to find critical lines, but what about when comparing nodes per seconds?

According to Vasik he says about the Rybka 2.2n2:
"This updated version of Rybka will display node counts which represent the multi-processing efficiency. (They are adjusted to take into account the wasted work.) So, you'll be able to compare the strength of different hardware, ie. you can compare a 4-core machine with a 2-core machine. The one with the better nodes-per-second is better "


Also i remember him saying that for next versions of Rybka 2.2n2 that(the aforementioned) will be the standard way of measuring speedup efficiency on multi-processors.

So nodes per second should give the true efficiency for Rybka!

So why the nodes per second does show only a 7.5 factor on a Skulltrail with 8 cores 4.8GHz Versus an Athlon 1 CPU 2.2GHz?

If we compare them:
•Athlon is 1 single CPU while Skulltrail is 8 CPUs(cores). So we have a speedup factor of 8x from this.
•But also this Athlon is 2.2GHz while the octal is 4.8 GHz. So we have a speedup factor of about 2x from this.

The "nodes per second" speedup(that Vasik said it should give the efficiency of Rybka) factor is 7.5 as i've said earlier.

If we now remove the second factor of speedup(the different GHz values) we then have:
Theoretical speedup: 1 CPU versus 8 CPUs gives 8x
Actual speedup based on "nodes per second": 1 CPU versus 8 CPUs for Rybka 3 gives a 7.5/2 = 3.75x

So Rybka's speedup from 1 to 8 CPUs seems to be for that position 3.75.
I don't understand why it's so low. What is going on wrong?

Personally that statement about comparing NPS is a crock of crap. You can take the same program and run it miltiple times on the same position and get different node counts each time. Sometimes _significantly_ different node counts.

Trying to adjust the node counts to reflect the SMP overhead is just silly, because there is no way to _measure_ SMP overhead on a position unless you do the following:

search it with one processor to the key depth and remember the nodes searched. Then re-search with all the processors used, to the same depth, and remember the nodes search. Compare the two. Usually the SMP search will search extra nodes and these are overhead. But it is _impossible_ to recognize overhead nodes in an SMP search because there is no way to know you are searching something the serial search would not search, unless you do the above.

Lots of nosense, little reality...

The question here is not about nodes per solution but "nodes" per second(we know that it is not nodes but something else that the program counts)

The question is if "nodes" was designed in a way that "nodes" per second is proportional to effective speed of the machine for deep rybka3.

And the answer is "it is not possible to do this". There are two kinds of overhead in a parallel search.

1. You start multiple threads at a position where you believe you have to search all branches. But then one fails high. All the branches searched there by other processors is wasted effort (99% of the time, in maybe 1% of the time one of these branches actually helps the fail high happen through transposition table sharing so this fraction is impossible to measure). So one _could_ claim all of this is overhead, and be fairly accurate.

2. Multiple search threads do their thing, but they can do it without having perfect information. For example, in the classic YBW approach, after searching the first move, you start a parallel search, believing that you now have the alpha (lower) bound value established. But once you start searching in parallel, alpha can be improved, and all the stuff searched by other processors using the old alpha value will have extra nodes that would not have been necessary had the true alpha value been known. There is no way to measure this, and no way to reliably repeat it either..

There are some questions that only people who have deep rybka3 can answer by comparing results on different machines.

1)Does Deep rybka report the same number of nodes per second if you run the same position twice(I do not care about difference that are not more than 5%)?

What would that prove? That he still mangles the node counter? Crafty is very consistent in its NPS given the same position. Because NPS is a function of search speed, and SMP overhead such as lock interference, and architectural issues that can vary from program to program (cache size, cache bandwidth, memory latency, processor scheduling and affinity, etc).

2)Is it correct that rybka plays better in machines with more nodes per second?

That is probably trivial to answer as most any program gets _some_ sort of speedup with SMP, even if they are poorly written. And any speedup will translate into better chess.

3)Is it correct that the number of nodes per second is proportional to the playing strength?

Again, true for most any program, when comparing the same program at two different NPS speeds..

In other words suppose that rybka shows 10000 nodes per second in machine A and 30,000 nodes per second in machine B

Does it mean that rybka is practically doing something that is equivalent to being 3 times faster in machine B?

Hard to say. One could mangle the NPS value to approximate the average SMP speedup. But it will be an "average" and vary wildly from the truth at times. Hence the idea is just another way to confuse, rather than enlighten users about what is going on.

[S.Taylor]

Very bad, Alburt! It's just not good enough. It's EVIL!

I have been testing Rybka 3, both the 32-bit and 64-bit engines, on a single CPU machine, testing in controlled matches as well as various test suites and the like.

Since Rybka 232a is still the undisputed number one engine, the announcement of an upgraded and significantly stronger version, with a long list of improvements and speculative Elo increase, has led to a lot of questions on how much is hype and how much is reality.

I don't have the final answer to this, but do believe what I know and have seen is close to the truth of it.

The first thing I did, after reading the touted tactical and kingside attack improvements, was to run it through some positions. I chose Michael Gurevich's 40 position suite, with a dozen king attack positions which had at least 3 positions that had completely stumped Rybka 232a on my rig, a modest Athlon64 2.2GHz with 512MB hash.

The result was impressive: Rybka 3 32-bit solved every single one of them and only once took as long as 4 minutes. The solution times varied from 2 times faster than Rybka 232a, to more than 10 times. I also checked to see if the 64-bit version still offered its known speedups, and the results showed a consistent 37% speedup on my rig. Note that I have been told that on machines with faster memory and the like, the speedup may even be greater, but know that 37% is what can be expected at the very least. Here are a couple of examples:

MGV34-K

12: SOLOVJOV - Lugovoj, St. Petersburg 2005

[D]1rb1qrk1/ppn1p1b1/3p3p/2pP2p1/4P1nQ/P1N2N1P/1PPBB1P1/R4RK1 w - - 0 1


Analysis by Rybka 232a: Not found in 20 mins

Analysis by Rybka 3 32-bit :

1.Qg3 Nf6 2.Qe1 Nd7 3.Rd1 Qg6 4.Bd3 Ne5 5.Nxe5
= (0.20) Depth: 12 00:00:25 694kN
1.Qg3 Nf6 2.Qe1 Nd7 3.Bd3 Ne5 4.Nxe5 Rxf1+ 5.Qxf1 Bxe5 6.Qf3 Qf7
= (0.21) Depth: 13 00:00:40 1123kN
1.Qg3 Nf6 2.Bd3 Nh5 3.Qe1 Nf4 4.Bxf4 Rxf4 5.a4 Bd7 6.Ne2 Rf8 7.Qa5
= (0.17) Depth: 14 00:01:26 2424kN
1.Qg3 Nf6 2.Bd3 Nh5 3.Qe1 Nf4 4.Bxf4 Rxf4 5.e5 c4 6.Be4 Nb5 7.Qe3 Nxc3 8.bxc3 Qd8 9.Rab1 dxe5 10.Nd2
= (0.23) Depth: 15 00:02:48 4860kN
1.Bxg5
+/= (0.37) Depth: 15 00:04:00 6741kN
1.Bxg5
+/= (0.57) Depth: 15 00:04:34 7489kN


Analysis by Rybka 3 64-bit :

1.Qg3 Nf6 2.Qe1 Nd7 3.Rd1 Qg6 4.Bd3 Ne5 5.Nxe5
= (0.20) Depth: 12 00:00:16 694kN
1.Qg3 Nf6 2.Qe1 Nd7 3.Bd3 Ne5 4.Nxe5 Rxf1+ 5.Qxf1 Bxe5 6.Qf3 Qf7
= (0.21) Depth: 13 00:00:25 1123kN
1.Qg3 Nf6 2.Bd3 Nh5 3.Qe1 Nf4 4.Bxf4 Rxf4 5.a4 Bd7 6.Ne2 Rf8 7.Qa5
= (0.17) Depth: 14 00:00:54 2424kN
1.Qg3 Nf6 2.Bd3 Nh5 3.Qe1 Nf4 4.Bxf4 Rxf4 5.e5 c4 6.Be4 Nb5 7.Qe3 Nxc3 8.bxc3 Qd8 9.Rab1 dxe5 10.Nd2
= (0.23) Depth: 15 00:01:46 4860kN
1.Bxg5
+/= (0.37) Depth: 15 00:02:32 6741kN
1.Bxg5
+/= (0.57) Depth: 15 00:02:53 7489kN

MGV39-K

14: CONQUEST - Hodgson, BL 9596 1995

[D]4kbnr/ppq2ppp/1np1p3/6P1/3r3P/2N2Q2/PPPB1P2/2KR1B1R w k - 0 1

Analysis by Rybka 232a: Not found in 20 mins

Analysis by Rybka 3 32-bit :

1.h5 Ne7 2.h6 Ned5 3.hxg7 Bxg7 4.Nxd5 cxd5 5.c3 a6 6.Kb1 Ra4
+/= (0.40) Depth: 13 00:00:35 974kN
1.h5 Ne7 2.h6 Ned5 3.hxg7 Bxg7 4.Nxd5 cxd5 5.c3 a6 6.Kb1 Ra4 7.b3 Ra5 8.Bd3 Rc5 9.Rh3 Nd7
+/= (0.35) Depth: 14 00:01:02 1747kN
1.Nb5
+/= (0.60) Depth: 14 00:01:40 3003kN
1.Nb5
+/- (0.80) Depth: 14 00:02:10 3995kN
1.Nb5
+/- (1.20) Depth: 14 00:02:19 4212kN
1.Nb5 cxb5 2.Bxb5+
+- (1.98) Depth: 14 00:03:24 6509kN
1.Nb5 cxb5 2.Bxb5+ Kd8 3.Bf4 Rd6 4.Rhe1 Kc8 5.Bxd6 Bxd6 6.Rxd6 Qxd6 7.Qc3+ Nc4 8.Bxc4 Ne7 9.Bxe6+ Kb8 10.Bxf7 Qf4+ 11.Qd2 Qc7 12.Qe3 Rc8 13.Qe5 Nc6 14.Qxc7+ Kxc7 15.Bd5 Kd6 16.Rd1
+- (2.10) Depth: 15 00:04:41 9081kN
1.Nb5 cxb5 2.Bxb5+ Kd8 3.Bf4 Rd6 4.Rhe1 Kc8 5.Bxd6 Bxd6 6.Rxd6 Qxd6 7.Qc3+ Nc4 8.Bxc4 Ne7 9.Bxe6+ Kb8 10.Bxf7 Qf4+ 11.Qd2 Qc7 12.Qe3 Rc8 13.Qe5 Nc6 14.Qxc7+ Kxc7 15.Bd5 Kd6 16.Rd1
+- (2.13) Depth: 16 00:07:17 14349kN
1.Nb5 cxb5 2.Bxb5+ Kd8 3.Bf4 Rd6 4.Rhe1 Kc8 5.Bxd6 Bxd6 6.Rxd6 Qxd6 7.Qc3+ Nc4 8.Bxc4 Ne7 9.Bxe6+ Kb8 10.Bxf7 Qf4+ 11.Qd2 Qc7 12.Qe3 Rc8 13.Qe5 Nc6 14.Qxc7+ Kxc7 15.Bd5 Kd6 16.Rd1
+- (2.13) Depth: 17 00:12:48 24568kN


Analysis by Rybka 3 64-bit:

1.h5 Ne7 2.h6 Ned5 3.hxg7 Bxg7 4.Nxd5 cxd5 5.c3 a6 6.Kb1 Ra4
+/= (0.40) Depth: 13 00:00:22 974kN
1.h5 Ne7 2.h6 Ned5 3.hxg7 Bxg7 4.Nxd5 cxd5 5.c3 a6 6.Kb1 Ra4 7.b3 Ra5 8.Bd3 Rc5 9.Rh3 Nd7
+/= (0.35) Depth: 14 00:00:39 1747kN
1.Nb5
+/= (0.60) Depth: 14 00:01:02 3003kN
1.Nb5
+/- (0.80) Depth: 14 00:01:22 3995kN
1.Nb5
+/- (1.20) Depth: 14 00:01:27 4212kN
1.Nb5 cxb5 2.Bxb5+
+- (1.98) Depth: 14 00:02:08 6509kN

I also tested it using the Silver Suite, an openings suite with 50 positions, tested with both white and black. I tested against Fruit 234n (thanks to Ryan Benitez) which had scored decently against Rybka 232a 32-bit:

Code: Select all

	
1	Rybka 2.3.2a 32-bit		+47/-13/=40	67.0/100
2	Fruit 2.3.4n 		      +13/-47/=40	33.0/100

Rybka 232a had a +123 Elo performance in this match-up. I expected Rybka 3 to improve on this, but figured the results we have seen for quads and octas and other latin-number configurations wouldn't be as extreme on a single-CPU machine. My, was I wrong....

Code: Select all

	
1	Rybka 3 32-bit		+71/-10/=19	80.5/100
2	Fruit 2.3.4n  		+10/-71/=19	19.5/100

Rybka 3 32-bit had a +246 elo performance, in other words a +123 Elo improvement over Rybka 232a 32-bit.

It's hard to know what to say, since the genuine numbers themselves appear straight out of a marketing blurb. The 64-bit single-CPU will truly be about 200 elo stronger than the next non-Rybka engine, which is just insane...

Albert

[Milton]

Albert,

If I may ask, was this result obtained with Rybka's default contempt settings? Thanks.

Milton

[Albert Silver]

Albert,

If I may ask, was this result obtained with Rybka's default contempt settings? Thanks.

Milton

Yes, I made no changes to the default settings of the engine.

[ernst]

Rybka 3 needs 19 ply and 10:48 on my Q6600 @ 3GHz and 2GB hash, see below. Of course MP programms are non-deterministic.

Code: Select all

New game - Rybka 3, Blitz:15'+10"
1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by Rybka 3:

1.Bxa7 
  +/-  (0.93 !)   Depth: 6   00:00:00  6kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4 
  +/-  (0.85)   Depth: 6   00:00:00  12kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4 
  +/-  (0.85)   Depth: 7   00:00:00  19kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4 Rd8 
  +/-  (0.80)   Depth: 8   00:00:00  45kN
1.Bxa7 
  +/-  (1.00 !)   Depth: 9   00:00:00  123kN
1.Bxa7 
  +/-  (1.20 !)   Depth: 9   00:00:00  145kN
1.Bxa7 
  +-  (1.60 !)   Depth: 9   00:00:01  172kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 
  +/-  (1.33)   Depth: 9   00:00:01  272kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 
  +/-  (1.35)   Depth: 10   00:00:01  340kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 7.Kg1 Be7 
  +/-  (1.24)   Depth: 11   00:00:02  512kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Rd7 6.Ra1 Nc6 7.Bh3 Ncd4 
  +/-  (1.20)   Depth: 12   00:00:04  1090kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Rd7 6.Ra1 Nc6 7.Bh3 Ncd4 
  +/-  (1.20)   Depth: 13   00:00:06  1620kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Rc8 6.Be3 Rd8 7.Qa5 Qxa5 
  +/-  (1.12)   Depth: 14   00:00:10  2650kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Qe3 Ne6 5.Bb6 Rd7 6.Bh3 Qc6 7.Bxe6 Qxe6 8.Ra1 Nc8 9.Kg2 Be7 
  +/-  (1.10)   Depth: 15   00:00:20  5682kN
1.Bxa7 Bxe4 2.dxe4[] e5 3.bxc5 Nxc5 4.Rd1 Rxd1+ 5.Qxd1 Nc8 6.Bb8 Na6 7.Bd6 Nb4 8.Bxb4 Qxb4 9.Kg1 
  +/-  (1.14)   Depth: 16   00:00:39  11496kN
1.Bxa7 Bxe4 2.dxe4[] e5[] 3.bxc5 Nxc5 4.Rd1 Rxd1+ 5.Qxd1 Nc8 6.Bb8 Na6 7.Bd6 Nb4 8.Bxb4 Qxb4 9.Kg1 b5 10.Ne3 Ne7 11.h4 Nc6 12.Nd5 Qd4 13.Bf3 Qc5 14.Qd2 
  +/-  (1.14)   Depth: 17   00:01:14  23264kN
1.Bxa7 Bxe4 2.dxe4[] e5[] 3.bxc5 Nxc5 4.Rd1 Rxd1+ 5.Qxd1 Nc8 6.Bb8 Na6 7.Bd6 Nb4 8.Bxb4 Qxb4 9.Kg1 b5 10.Ne3 Ne7 11.h4 Nc6 12.Nd5 Qd4 13.Bf3 Qc5 14.Qd2 
  +/-  (1.14)   Depth: 18   00:02:28  48244kN
1.Bxa7 Bxe4 2.dxe4[] e5[] 3.bxc5 Nxc5 4.Rd1 Rxd1+ 5.Qxd1 Nc8 6.Bb8 Na6 7.Bd6 Nb4 8.Bxb4 Qxb4 9.Kg1 b5 10.Ne3 Ne7 11.h4 Qd4 12.Qxd4 exd4 13.Ng4 Nc6 14.f4 d3 15.cxd3 Bd4+ 16.Kf1 
  +/-  (1.14)   Depth: 19   00:05:53  113mN
1.Qxf6+ 
  +/-  (1.34 !)   Depth: 19   00:10:48  213mN
1.Qxf6+ 
  +-  (1.54 !)   Depth: 19   00:11:05  216mN
1.Qxf6+ 
  +-  (1.94 !)   Depth: 19   00:11:33  221mN
1.Qxf6+ 
  +-  (2.74 !)   Depth: 19   00:12:24  231mN

(,  29.07.2008)

These positions were from Rybka 3 on 8 core Skulltrail at 4.8Ghz

New game, Blitz:16'
[d]1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by Rybka 3 :

1.Bxa7
² (0.58) Depth: 4 00:00:00
1.Bxa7
² (0.58) Depth: 4 00:00:00
1.Bxa7
² (0.58) Depth: 4 00:00:00
1.Bxa7 Bxe4
± (0.73) Depth: 5 00:00:00
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4
± (0.85) Depth: 6 00:00:00 17kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4
± (0.85) Depth: 6 00:00:00 17kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4
± (0.85) Depth: 7 00:00:00 25kN
1.Bxa7 Bxe4 2.Qxe4 Nc3 3.Qxb7 Qxb7 4.Bxb7 Rd7 5.Na5 cxb4 6.Bc5 Ned5 7.Bxd5 Nxd5 8.Nc4 Rd8
± (0.79) Depth: 8 00:00:00 56kN
1.Bxa7
± (0.99) Depth: 9 00:00:00 119kN
1.Bxa7
± (1.19) Depth: 9 00:00:00 125kN
1.Bxa7
+- (1.59) Depth: 9 00:00:00 138kN
1.Bxa7 Bxe4 2.dxe4 e5 3.Ne3 Nb6
± (1.35) Depth: 9 00:00:00 219kN
1.Bxa7 Bxe4 2.dxe4 e5 3.Ne3 Nb6 4.Ng4 Ng8 5.bxc5 Qxc5 6.Ra1 Rd2 7.Ne3 Bd8
± (1.35) Depth: 10 00:00:00 242kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Ra8 6.Qd2 Nd4
± (1.13) Depth: 11 00:00:01 605kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Ra8 6.Ra1 Rxa1+ 7.Qxa1 Nc6 8.Qa8

± (1.16) Depth: 12 00:00:01 788kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Rd1 Ra8 5.Bxc5 Qxc5 6.Rd6 Ng8 7.Rd5 Qb4 8.Kg1 Ra1+ 9.Bf1 Be7
± (1.24) Depth: 13 00:00:03 1435kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 7.h4 Nd4 8.Ba5 Qb5 9.Kg1 Qc6 10.Bb6
± (1.20) Depth: 14 00:00:04 2459kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 7.h4 Nd4 8.Ba5 Qb5 9.Kh2 h6 10.Bc3 Nce6 11.Bd2
± (1.24) Depth: 15 00:00:08 4738kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 7.h4 Nd4 8.Ba5 Qb5 9.Kh2 h6 10.Bc3 Nce6 11.Bd2
± (1.24) Depth: 16 00:00:16 9083kN
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Rd1 Qb4 5.Bb6 Rxd1+ 6.Qxd1 Nc6 7.h4 Be7 8.Kh2 Qc3 9.Ne3 Bf8 10.Nd5 Qb2
± (1.16) Depth: 17 00:00:35 20688kN
1.Qxf6+
+- (1.44) Depth: 17 00:01:00 36506kN
1.Qxf6+
+- (1.64) Depth: 17 00:01:05 38645kN
1.Qxf6+
+- (2.04) Depth: 17 00:01:20 44783kN
1.Qxf6+
+- (2.84) Depth: 17 00:01:34 50095kN

(, 21.01.2006)