Rene, On Tue, Jun 22, 2010 at 9:58 AM, René van de Veerdonk < [email protected]> wrote:
> Fuming and Petr, > > I was giving the tree-search some thought over the weekend and made some > crude estimates based on my own MCTS implementation. I did not make fresh > timing measurements, just added/subtracted some numbers to get to where I > want to be. Assuming you want to maintain RAVE and win-rate information, I > am guesstimating that for my quasi-heavy playouts I find an overhead of > about 17% for the tree-search part (50,000 playouts total). My quasi-heavy > playouts run at 20,000 playouts/second, and I am estimating that if you > appropriately break the tree-search portion in descend, update, copy, and > misc threads, you can get the overhead to come down to, say, 5%. Then, if > you completely outsource the playouts to the FPGA, i.e., you make them > cost-free (ignoring communication costs), you would be able to boost the > playouts twenty-fold, to 400,000, at which point you become CPU limited. > Assume further that my coding can be further improved by another 2x (not > unreasonable at all), and you get to 800,000. > I it is feasible to send playouts out to the FPGA, because playouts are dynamic, depending on the structure of the tree, therefore, FPGA has to run simulations on small number of playouts very fast, which FPGA is not good at, It is good at running playouts of one board position for thousands of times quickly (at least our implementation is like this). > Fuming, are you planning to run a version of your program on KGS/CGOS at > some time in the future? Than we can all see it in action. > We have not such plans. Right now, the system requires human intervention to start and end a game. We went to a national computer go event in Shenzhen, China, but performed poorly. We have made improvements since, maybe it will do better this year. Best, Fuming
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
