Zax,

CUDA is applicable to a huge variety of compute problems, much more so than the 
old GPGPU linear algebra canonical examples (see www.nvidia.com/cuda for a 
large list of problems and their speedups from using CUDA).  Quackle seems like 
it should be easily parallelizable, since each simulation iteration could be a 
thread, but the threads would require a ton of branching and lookups into data 
structures that probably don't map well to GPUs.  If the inner loop has few 
conditionals and a lot of floating point math, CUDA will work well and isn't 
terribly hard to implement.  I haven't used CUDA much, but I suspect move 
generation alone would be incredibly hard to implement.  I think having to keep 
indepenent lists of moves and sort them, and update the board, etc, pretty much 
puts the nail in the head.

If you could find a clever mapping of Scrabble to Navier-Stokes, though, then 
we're gold.  :)

Eugene d'Eon
NVIDIA



________________________________
From: "[email protected]" <[email protected]>
To: [email protected]
Sent: Wednesday, April 8, 2009 7:22:02 AM
Subject: [quackle] GPU acceleration for Quackle.



Anyone familiar with NVIDIA Cuda?  Seems like a crazy way to accelerate 
simulation.

Reply via email to