Houdini 3 the same in 64-bit and 32-bit: well, not quite???

[ernest]

Found this example in some "other" Forum,
and tested it, to large depths, with my no-POPCNT Houdini 3 update version.
Maybe somebody can confirm on a modern configuration (i7...)

Until depth 21: no difference in nodes
Depth 22 23 24: very small difference in nodes
Depth 25: large difference in nodes
Depth 26 and more: analysis is different

Infinite analysis
16 MB hash 1 thread
[d]r1qr1bk1/pppbnppp/4n3/3NP3/8/5N1P/PB2QPP1/1BRR2K1 w - -
Engine: Houdini 3 w32 by Robert Houdart
16 MB Hash
1 thread used
16/50 0:05 +1.53 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Be7 7.Qe3 Bc5 8.Nd4 a6 9.Rd2 Bb5 10.a3 Bc6 11.f3 Bd5 12.h4 Bc4 (8.884.794) 1668
17/50 0:11 +1.65++ 1.Bxh7+ (20.523.585) 1746
17/50 0:12 +1.57 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Bc6 7.Rxd8+ Rxd8 8.Bc1 Rd5 9.Be3 a6 10.Kh2 Be7 11.h4 Rb5 12.g3 Be4 13.a4 Rd5 (22.255.807) 1748
18/50 0:22 +1.57 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Bc6 7.Rxd8+ Rxd8 8.Bc1 Rd5 9.Be3 a6 10.Kh2 Be7 11.h4 Rb5 12.g3 Be4 13.a4 Rd5 (41.243.077) 1801
19/50 0:30 +1.69++ 1.Bxh7+ (54.261.487) 1791
19/50 0:34 +1.66 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Qe3 Rc5 6.Rd2 Ra5 7.Nd4 Be7 8.Nxe6 Bxe6 9.Rxd8+ Bxd8 10.Qe4 Rd5 11.Bd4 b6 12.a3 Be7 13.a4 Rd8 14.g4 (61.956.869) 1793
20/52 1:17 +1.78++ 1.Bxh7+ (141.111.349) 1814
20/52 1:29 +1.72 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Qe3 Rc5 6.Rd2 Ra5 7.Nd4 Be7 8.Qf3 Nxd4 9.Rxd4 Bc6 10.Rxd8+ Bxd8 11.Qd3 Rd5 12.Qe2 Bd7 13.Kh2 Be6 14.f4 (162.998.385) 1824
21/56 2:46 +1.84++ 1.Bxh7+ (306.418.277) 1835
21/56 3:31 +1.81 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qb1 Rxd1 9.Qxd1 Rd8 10.Qb1 Nc6 11.h4 Bd7 12.h5 Be7 13.Qf1 a6 14.Nf5 (388.169.197) 1836
22/61 7:12 +1.81 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qb1 Rxd1 9.Qxd1 Rd8 10.Qb1 Nc6 11.h4 Bd7 12.h5 Be7 13.Qf1 a6 14.Nf5 (822.026.966) 1899
23/61 12:31 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (1.433.989.025) 1906
24/68 28:03 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (3.300.559.647) 1960
25/68 58:32 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (7.021.444.358) 1998
26/68 107:32 +2.03++ 1.Bxh7+ (12.986.739.310) 2012
26/68 127:54 +2.28++ 1.Bxh7+ (15.198.504.350) 1980
26/72 188:22 +2.19 1.Bxh7+ Kxh7 2.Rxc7 Ng6 3.Rxc8 Raxc8 4.Nc3 Ngf4 5.Qe4+ Kg8 6.Ng5 Nxg5 7.Qxf4 Ne6 8.Qg4 Bc6 9.Rxd8 Rxd8 10.Ne4 Be7 11.Nf6+ Kf8 12.Qh4 Bc5 13.Ne4 Bxe4 14.Qxe4 (22.647.255.461) 2003

Engine: Houdini 3 x64 by Robert Houdart
16 MB Hash
1 thread used
16/50 0:04 +1.53 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Be7 7.Qe3 Bc5 8.Nd4 a6 9.Rd2 Bb5 10.a3 Bc6 11.f3 Bd5 12.h4 Bc4 (8.884.794) 2115
17/50 0:09 +1.65++ 1.Bxh7+ (20.523.585) 2197
17/50 0:10 +1.57 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Bc6 7.Rxd8+ Rxd8 8.Bc1 Rd5 9.Be3 a6 10.Kh2 Be7 11.h4 Rb5 12.g3 Be4 13.a4 Rd5 (22.255.807) 2199
18/50 0:18 +1.57 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Rd2 Bb4 6.Rd1 Bc6 7.Rxd8+ Rxd8 8.Bc1 Rd5 9.Be3 a6 10.Kh2 Be7 11.h4 Rb5 12.g3 Be4 13.a4 Rd5 (41.243.077) 2251
19/50 0:24 +1.69++ 1.Bxh7+ (54.261.487) 2240
19/50 0:27 +1.66 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Qe3 Rc5 6.Rd2 Ra5 7.Nd4 Be7 8.Nxe6 Bxe6 9.Rxd8+ Bxd8 10.Qe4 Rd5 11.Bd4 b6 12.a3 Be7 13.a4 Rd8 14.g4 (61.956.869) 2240
20/52 1:02 +1.78++ 1.Bxh7+ (141.111.349) 2264
20/52 1:11 +1.72 1.Bxh7+ Kxh7 2.Rxc7 Nxd5 3.Rxc8 Raxc8 4.Rxd5 Kg8 5.Qe3 Rc5 6.Rd2 Ra5 7.Nd4 Be7 8.Qf3 Nxd4 9.Rxd4 Bc6 10.Rxd8+ Bxd8 11.Qd3 Rd5 12.Qe2 Bd7 13.Kh2 Be6 14.f4 (162.998.385) 2276
21/56 2:14 +1.84++ 1.Bxh7+ (306.418.277) 2285
21/56 2:49 +1.81 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qb1 Rxd1 9.Qxd1 Rd8 10.Qb1 Nc6 11.h4 Bd7 12.h5 Be7 13.Qf1 a6 14.Nf5 (388.169.197) 2285
22/61 5:47 +1.81 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qb1 Rxd1 9.Qxd1 Rd8 10.Qb1 Nc6 11.h4 Bd7 12.h5 Be7 13.Qf1 a6 14.Nf5 (822.026.943) 2363
23/61 10:05 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (1.433.988.989) 2368
24/68 22:40 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (3.300.559.605) 2425
25/68 48:37 +1.91 1.Bxh7+ Kxh7 2.Rxc7 Kg8 3.Rxc8 Raxc8 4.Qe3 b5 5.Kh2 Nf5 6.Qd3 Ne7 7.Ne3 Be8 8.Qe2 Rxd1 9.Nxd1 b4 10.Ne3 Rd8 11.Qa6 Nc6 12.Qc4 Bc5 13.h4 Bd7 14.h5 (7.248.502.874) 2484
26/68 81:16 +2.03++ 1.Bxh7+ (12.055.153.391) 2472 12.986.739.310
26/77 121:59 +2.06 1.Bxh7+ Kxh7 2.Rxc7 Ng6 3.Rxc8 Raxc8 4.Nc3 Ngf4 5.Qe4+ Kg8 6.Ng5 Nxg5 7.Qxf4 Ne6 8.Qg4 Bc6 9.Rxd8 Rxd8 10.Ne4 Be7 11.Nf6+ Kf8 12.Qh4 Bc5 13.Ne4 Bxe4 14.Qxe4 (18.081.045.916) 2470
27/77 213:10 +2.18++ 1.Bxh7+ (32.230.645.162) 2519
27/77 275:30 +2.13 1.Bxh7+ Kxh7 2.Rxc7 Ng6 3.Rxc8 Raxc8 4.Nc3 Ngf4 5.Qe4+ Kg8 6.Ng5 Nxg5 7.Qxf4 Ne6 8.Qg4 Bc6 9.Rxd8 Rxd8 10.Ne4 Be7 11.Nf6+ Kf8 12.Qh4 Bc5 13.Ne4 Bxe4 14.Qxe4 (42.073.619.039) 2545

[Houdini]

Crazy - first difference after 800 million nodes .
I'll investigate, thanks for reporting.

[Houdini]

After investigation it appears that it's (again) the 32-bit version that doesn't behave as expected.

In positions without any pawns, the piece value for the Knight will be slightly inaccurate (0.36 pawns to be exact).
For example in the following position which is the exact position that gave rise to the analysis difference posted above:

[D]3B4/R7/6k1/8/8/7b/n4K2/8 b - -
The impact on playing strength is insignificant - in position without pawns it doesn't really matter whether the knight is evaluated as 3.1 or 2.8 pawns.
We'll simply include the correction for this issue in the next release.

Thanks for the bug report!

Robert

[ernest]

Thanks for the bug report!

Thanks for the quick analysis!
Indeed, unsignificant bug...

[ernest]

In positions without any pawns, the piece value for the Knight will be slightly inaccurate (0.36 pawns to be exact).

Hi Robert,

I just found that for this position with pawns
(studied elsewhere in the Forum,
http://www.talkchess.com/forum/viewtopi ... 257#498257
solution Nc6!),

[d]1N4K1/1pp1p3/7k/2r3pP/p2pp1P1/P6N/8/2B5 w - - 0 1
the 1-thread infinite analysis of the 32 and 64-bit versions differ as early as depth 16 and 3369 kN (I used 1024 MB hash).
32-bit:
1.Nd7 Rxc1
-+ (-2.29 --) Depth: 16/57 00:00:01 1843kN
1.Nd7
-+ (-1.58 !) Depth: 16/58 00:00:02 3369kN
64-bit:
1.Nd7 Rxc1
-+ (-2.29 --) Depth: 16/57 00:00:00 1843kN
1.Nd7
-+ (-1.58 !) Depth: 16/61 00:00:01 3673kN

[Houdini]

Hello Ernest, this is the same error as above.
On my current 32-bit version with correction I find the same output as the 64-bit version.
Thanks for taking the time to report this!

Robert

[Dr.Wael Deeb]

Hello Ernest, this is the same error as above.
On my current 32-bit version with correction I find the same output as the 64-bit version.
Thanks for taking the time to report this!

Robert

Hope you'll be able to fix it this time......
Dr.D

[kranium]

Hello Ernest, this is the same error as above.
On my current 32-bit version with correction I find the same output as the 64-bit version.
Thanks for taking the time to report this!

Robert

Hope you'll be able to fix it this time......
Dr.D

he might not have time...
seems far too busy shamelessly singing Houdini's glory at every possible opportunity!


that being said...
i do wish you and all of CCC a very Happy New Year!


Best Wishes,
Norm

[Dr.Wael Deeb]

Hello Ernest, this is the same error as above.
On my current 32-bit version with correction I find the same output as the 64-bit version.
Thanks for taking the time to report this!

Robert

Hope you'll be able to fix it this time......
Dr.D

he might not have time...
seems far too busy shamelessly singing Houdini's glory at every possible opportunity!


that being said...
i do wish you and all of CCC a very Happy New Year!


Best Wishes,
Norm

Happy new year Norman

I'm still waiting for a new chess release from your side......

To me,the FireBird chess engine is a masterpiece,particulary v1.1....
I have a personality of this engine which is extremely strong and is palying regulary in my tournaments.....
Dr.D

[Houdini]

Hope you'll be able to fix it this time......
Dr.D

The issue is perfectly irrelevant, as explained above it doesn't have any impact on the analysis quality.