Re: computer-go: FPGA vs. Clusters

[Vincent Diepeveen <diep@xxxxxxxxx>]

For local searches at small areas hashtable works great.
Communication at clusters is dead slow. In the computerchess world
we have huge problems with that.

For a 64 processor cluster with 1 usec latency and a program
searching around 20000 nodes a second i calculated the speedup
would be less as the square root, where the speedup of my program
at a dual machine with shared memory is 2.0 (1.9 at less as 2 
minutes a move) and at a quad it's real near to 4.0 at slower 
time controls. 

If you add FGPA you will suffer from more problems, like how to
do any communication between the fpga and a general purpose processor?

Apart from that FPGA speeds up the nodes a second.

Matrix calculations and several similar models are ideal for
clusters to give a great speedup. However game tree search is
very dependant upon the other searches. You can't split this
in independant searches in an easy way. Communication is vital
to get better speedups. In future when go programs will look ahead
deeper this problem will get real huge. Right now i can search
only 7 ply after the first stones are on the board, which is real bad.

At 02:23 PM 8/29/00 +0200, you wrote:
>> FPGA in combination with running at a cluster will not speed your
>> local tactical search up much. Cluster communication is pretty slow
>> compared to the speed of the local tactical search.
>
>Since this is local tactical search, there has to be several pending 
>tactical searches. If every processor in the cluster get the 
>responsibility for one tactical search at a time, then perhaps the 
>search times are long enough to make communication times negligeble. 
>It certainly has to be so in this case where the search has the speed 
>of 20 nodes a second, unless the the search terminates at 1 ply 
>because of the massive amount of knowledge used (that explains the 
>speed).
>
>By the way, I thought my program was slow which makes 20 down to .25 
>nodes per second when it makes whole board evaluations - including 
>all necessary reevaluations of  local tactical searches (tactical 
>searches do 2000-5000 nodes per second). 
>
>Magnus Persson
>--
>Magnus Persson
>Department of psychology, Uppsala University
>Box 1225, SE-751 42, Sweden
>Tel 018-471 2141 (work), 018-460264 (home)
>MAILTO: magnus.persson@xxxxxxxxxxxxxxxxx
>URL: http://www.docs.uu.se/~magnuspe
>
>
>