Hi all

I’m eager to get involved in the world of GPU-computing but need some advice before I take the plunge. I have done a little bit of reading but remain unsure about a few issues.

At the moment I am working on a fairly large academic project, part of which entails developing/implementing a genetic-algorithm to solve a nasty optimisation problem. My initial hope is to be able to exploit the power of GPU computing to speed up the genetic algorithm.

Without going into any concrete detail, one of the key characteristics of my genetic algorithm is that it runs for very many iterations where, at each iteration, there are many (let’s say no more than 200) sets of mathematically identical, CPU-intensive (e.g. involving difficult function evaluations) but independent calculations that need to be carried out. In addition to this, at the end of each iteration, some fairly minor processing needs to be applied to the combined results from all the independent calculations. Therefore the bulk of the algorithm seems amenable to a parallel implementation. I am using MATLAB.

My questions:

i.) if I use a card with, say, “200 CUDA cores”, does this mean in practical terms that I’ll readily be able to exploit each core to carry out a set of the aforesaid independent calculations at each iteration (suggesting an approx. 100-fold speedup in the algorithm), or is this wishful thinking?

ii.) my budget for the project will probably allow me to purchase a Tesla card - would this be worthwhile compared to buying, say, something like the GTX 580? The faster I can carry out each set of (in principle, fully parallel/independent) calculations, the better, but I’ll not need to carry out more than about 200 sets of calculations simultaneously. As for the amount of data involved in each independent calculation - probably several kb at most.

iii.) MATLAB’s Parallel Computing Toolbox doesn’t seem to support some of the functions I’ll want to run on the GPU (e.g. eig, if etc.) - is it easy to write my own overloaded functions or should I use Jacket, which seems to support rather more MATLAB functions?

iv.) any general comments on whether either Jacket or MATLAB’s Parallel Computing Toolbox would be better suited to my work?

I’d really appreciate any comments as I am totally new to the world of GPU computing. Thanks in advance.