Re: [Swarm-Modelling] ABMs on Graphical Processor Units

[Marcus G. Daniels]

..For some things (like CA type stuff), it would totally rock.  NVidia's 
Tesla systems just have ridiculous performance `in theory'.  But for 
general ABM, the Cell has a more general purpose instruction set and 
general purpose compilers (GCC instead of proprietary C-like subset 
compiler).   The GPUs, of course, lack high performance double precision 
arithmetic while the next generation of Cells have double precision.  
Java probably would be hard to make work well, because of the giant 
cache-busting footprint (and the small local store on the SPUs), but 
Objective C could message over the EIB in a fairly straightforward 
way.   The main thing would be to make everything a functional as 
possible (e.g. never touch memory space directly, but through methods) 
and introduce some spaces that distributed data over the local stores 
(this might not be that useful in practice, but of course a cool 
test-case is needed to show peak performance!).   To be useful, an 
implementation would need lots and lots of instrumentation to help the 
user in load balancing and to expose the reality of bottlenecks in the 
architecture..