Zeta Open CL Chess - v0.918 - Developer Release

[smatovic]

fyi:

Got AlphaBeta Pruning with Move Ordering running on one SIMD Unit of a GPU with 128 threads.

More Information:
http://zeta-chess.blogspot.com/

Source Code:
https://github.com/smatovic/Zeta/tree/zeta_nvidia_0918

--
Srdja

[JuLieN]

fyi:

Got AlphaBeta Pruning with Move Ordering running on one SIMD Unit of a GPU with 128 threads.

More Information:
http://zeta-chess.blogspot.com/

Source Code:
https://github.com/smatovic/Zeta/tree/zeta_nvidia_0918

--
Srdja

Wow, congrats, Srdja! How fast is it? (average nps)

[smatovic]

Wow, congrats, Srdja! Very Happy How fast is it? (average nps)

Thanks

one SIMD Unit with 128 threads makes 10 000 to 50 000 nps on my NV GTS250 with 16 SIMD Units on board,
but it is a Prototype, no castle moves or en passant inside...
a lot of improvements possible...

I am just glad that AB Pruning works also on the GPU

Next topic would be a load-balancer across mutlible SIMD Units.

--
Srdja

[JuLieN]

Hmm... this is quite slow, and as Intel's CEO said last week that Intel is going toward massively parallel CPUs I guess that GPUs will never catch up with CPUs.

BUT, IIRC, Apple designed OpenCL to make use of both CPU & GPU transparently, so your work is paving the future anyway

EDIT: yes, OpenCL handles both CPUs and GPUs. And both X86 and ARM CPUs will soon run OpenCL code!
http://semiaccurate.com/2011/04/26/arm- ... -and-gpus/

[smatovic]

Hmm... this is quite slow, and as Intel's CEO said last week that Intel is going toward massively parallel CPUs I guess that GPUs will never catch up with CPUs.

Yepp,
i dont think that a GPU Chess Programm will compete with a professional CPU based programm, there are too much restrictions in coding/design. Transposition tables, null move, iterative deeping, all this has to be reinvented for the GPU.

--
Srdja

[Albert Silver]

Hmm... this is quite slow, and as Intel's CEO said last week that Intel is going toward massively parallel CPUs I guess that GPUs will never catch up with CPUs.

BUT, IIRC, Apple designed OpenCL to make use of both CPU & GPU transparently, so your work is paving the future anyway

EDIT: yes, OpenCL handles both CPUs and GPUs. And both X86 and ARM CPUs will soon run OpenCL code!
http://semiaccurate.com/2011/04/26/arm- ... -and-gpus/

Well, bear in mind the GTS250 is quite slow overall compared to a top gaming card, and is much like a Celeron compared to a Sandy Bridge quad.

[smatovic]

Well, bear in mind the GTS250 is quite slow overall compared to a top gaming card, and is much like a Celeron compared to a Sandy Bridge quad.

Yepp,
the ATI 69xx series for example has x8 more processing elements per SIMD Unit but runs with half of clock rate, so a speedup of x4 could be possible.

I figured out that move ordering is the performance killer in Zeta,
somehow i have to design it more SIMD friendly.

--
Srdja

[voyagerOne]

What will the NPS be without move ordering?

[voyagerOne]

Wait you use a bubble sort???

[smatovic]

What will the NPS be without move ordering?

about 100 000 nps.

--
Srdja