Spark 0.3a (mp) released

[Allard Siemelink]

I noticed that hashfull displays value/threads.

Thanks for reporting, I'll see to it that this gets cleaned up in the next release.

[Allard Siemelink]

Spark 0.3a is now available from:

http://members.ziggo.nl/allard.siemelink/spark/

- MP support added for all platforms (Windows 32, Windows 64, Mac OSX 64), Linux 64 version will follow later
- bug fixes (eval, occaisional lockups)

Funny that the vc mp version for 64 bits is 10x bigger than all the others. Perhaps it contains debug symbols?

Turns out that it contains nearly 3Mb of zeros.
Somehow vc++ inserts these when compiling for SMP.
(it's probably just reserved space for the smp threads)

[Allard Siemelink]

Hi Allard,

I am pleasantly surprised by Spark. As example, a move from Rybka which is difficult for most engines.

best move: Qf3xf4 time: 5:53.655 min n/s: 8.867.094 nodes: 3.135.892.328

Wow, pretty good nps too. What hardware are you running on?

Nothing special these days, Q6600 @ 3.24GHz, using spark-0.3a-vc64-mp.exe

Ah, I see. You're getting ~2200 knps per core, which is double of what I am used to seeing (on a 32-bit opteron@2500).

[Allard Siemelink]

Using 8 threads on my Mac Pro:

Code: Select all

info depth 22 seldepth 42 score cp 207 time 55216 nodes 858721291 nps 15552037 hashfull 125 pv f3f4 e5f4 e1e6 c6e5 e6e5 f6e5 d1d7 c7b8 b3c1 f8f5 c1d3 b7b6 d7f7 b6c5 f7f5 b8b2 d3b2 g6f5 b2d3 e5e4 d3f4 a8b8 c4e6 b8b1 g1h2 b1b2 e6f5 b2c2

Not much different than your quad.

15500 knps is almost double of what Ernst is getting.
The ~2000knps/core is indeed comparable, thanks.

[Allard Siemelink]

Thanks. - Do you have a Spark 0.3 result for comparison, from the same conditions (except single/dual)?

Might just be bad luck over a small number of games so far, but I am getting worse results for Spark 0.3a than I had for Spark 0.3.

I just played 4000 games spark 0.3a vs spark 0.3. The elo difference was exactly 0.
The games were played with extremely fast time control, though: 4096 nodes/move (~3 games/second)

[Damir]

Hi Allard,

I am pleasantly surprised by Spark. As example, a move from Rybka which is difficult for most engines.

best move: Qf3xf4 time: 5:53.655 min n/s: 8.867.094 nodes: 3.135.892.328

Wow, pretty good nps too. What hardware are you running on?

Nothing special these days, Q6600 @ 3.24GHz, using spark-0.3a-vc64-mp.exe

Ah, I see. You're getting ~2200 knps per core, which is double of what I am used to seeing (on a 32-bit opteron@2500).

Hi Allard

Such a great programmer like yourself, on such an obsolete hardware.
you deserve better than that. Why not switch to 64 bits ? 32 bits belong to the age of dino's...
Regards
Damir

[Allard Siemelink]

Ah, I see. You're getting ~2200 knps per core, which is double of what I am used to seeing (on a 32-bit opteron@2500).

Hi Allard

Such a great programmer like yourself, on such an obsolete hardware.
you deserve better than that. Why not switch to 64 bits ? 32 bits belong to the age of dino's...
Regards
Damir

Yeah, I agree. Future versions will be developed on the new system I recently bought (64 bit/core i5 750).

[Allard Siemelink]

Thanks. - Do you have a Spark 0.3 result for comparison, from the same conditions (except single/dual)?

Might just be bad luck over a small number of games so far, but I am getting worse results for Spark 0.3a than I had for Spark 0.3.

I just played 4000 games spark 0.3a vs spark 0.3. The elo difference was exactly 0.
The games were played with extremely fast time control, though: 4096 nodes/move (~3 games/second)

Made another run: +2 elo for 0.3a (4000 games at 8192 nodes/move, average depth 6.5 ply)

[Allard Siemelink]

Thanks. - Do you have a Spark 0.3 result for comparison, from the same conditions (except single/dual)?

Might just be bad luck over a small number of games so far, but I am getting worse results for Spark 0.3a than I had for Spark 0.3.

I just played 4000 games spark 0.3a vs spark 0.3. The elo difference was exactly 0.
The games were played with extremely fast time control, though: 4096 nodes/move (~3 games/second)

Made another run: +2 elo for 0.3a (4000 games at 8192 nodes/move, average depth 6.5 ply)

And finally: +7 elo for 0.3a over 2000 games at 32768 nodes/move, average depth 8.5 ply

If anything, these numbers suggest that Spark 0.3a is at least as strong as its predecessor.
There may be a trend that the improvement increases with longer time controls (making it ~20? elo stronger in Blitz )