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

[Roshan D'Souza]

FPGAs have limited on chip memory and very low off-chip memory bandwidth as compared to GDDR3/4/5. Also, since the code is directly converted to hardware, I am not sure if the real-estate on the FPGAs will be enough for even simple ABMs.

 

 

 

 

On 12/28/07, Marcus G. Daniels <address@hidden> wrote:

Russell Standish wrote:
> It sounds like one could do MPI on the Cell, however, which would allow
> Graphcode-based applications written in EcoLab to work.
Unfortunately, the ELF text section of OpenMPI is around 5 times too big
to fit on a Cell local store.   It might be possible to strip it down,
but after all that work you'd probably have something that looked a lot
like libspe2:

http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/1DFEF31B3211112587257242007883F3/$file/SPE_Runtime_Management_API_v2.2.pdf
>
> GPUs are a different kettle of fish, as they are definitely SIMD based
> (apply the same instruction to an array of pixels).
>
Nonetheless..

http://developer.nvidia.com/object/cuda.html
http://developer.download.nvidia.com/compute/cuda/1_1/NVIDIA_CUDA_Programming_Guide_1.1.pdf

It's also worth noting that development kits for field programmable gate
arrays  (FPGA) also have C dialect languages.

http://en.wikipedia.org/wiki/Impulse_C

For those times that you really want to be able to optimize over the
parameter space for an ABM.  :-)
_______________________________________________
Modelling mailing list
address@hidden
http://www.swarm.org/mailman/listinfo/modelling