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

[Marcus G. Daniels]

Roshan D'Souza wrot
> Cell processors are very limited in the sense that they do not allow
> dynamic branching in the code. 
Not sure what you mean by this.  It's true that Cell SPU processors lack
elaborate out of order execution units, but prefetch is possible.  For
example, objects coming up on a discrete event schedule could be
prefetched.

> Also, the memory bandwidth is lower as compared
> to the GPUs. 
I think the Cell is attractive because it is in many ways a general
purpose architecture with general purpose, open source tools.  Both
latency (5 nanoseconds) and bandwidth (~200 GB/s) are quite good
compared to conventional processors.  Working out ways to cache
subspaces back and forth across the EIB to limited local store could be
a challenge, but it is still very fast compared to what one would expect
from, say, a SMP Opteron system that would have inter-processor
latencies more on the order of 100 nanoseconds (or high performance
Infiniband/MPI of 1000 nanoseconds).    And while the local store is
very limited (256k), it's latency a lot like L2 cache on general purpose
CPUs.

Marcus