I have read about some colleagues running market simulations on them due to
their exceptionally high multithreading and fast memory, but I wondered if
there was some obvious limitation that I had overlooked before I start working
in this direction and learning a while new low level API.
[...]
Overally, there might be a bonus, but it's small and the coding and
debugging is much more laborous. Even when you have basic simulation
going, you need also tree descent. Then you need to start transferring
all the crucial heuristics to your GPU code, adding even more divergence.
But there has not been THAT much investigation, just the attempt of
Christian Nentwich and me. So maybe you would figure out some smart
tricks. :-)
Hi,
I've been working on this for a few months. So far I don't have any code
running on the GPU, but I expect to achieve performances in the 5 to 10
M playouts per second range on my current board (HD 5770, not a high end
one), for 9x9 with light playouts. No tree search.
When I started to think about this (after a discussion with a friend
about GPGPU potential), I decided to implement the algorithm on a CPU
first (better debugging environnement) but taking into account GPU
constraints. This is now completly tested, using SIMD instructions, and
I was amazed to discover it runs 1.6 faster than the libego light
playouts implementation (with extra bonuses such as generating moves
with true uniform distribution, as far as the underlying RNG is) on a
Q9650 under Linux.
I will start to port it to GPU soon. Just need some time, as it is a
night time hobby...
Regards,
Antoine
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
