In general, it is quite easy to get started on GPU computations in Matlab, what you have to do is to cast your variables to the gpuArray class: gvar=gpuArray(var);

If you can live with single precision (32-bit floating) point you should cast your variables to singles first: var=single(var) as that will give you a lot larger speedup in particular for consumer grade cards like the 1060. Afterwards working on these arrays will use the GPU (note that not all functions will be supported). As to how you get speedup it depends a lot on the specific problem is but where you in general archive large speedups is for batched operations on reasonable large arrays. For these purposes in particular the functions bsxfun (https://www.mathworks.com/help/matlab/ref/bsxfun.html) and pagefun (https://uk.mathworks.com/help/parallel-computing/pagefun.html) are your friends. In your setting I’m guessing that the high cost operation might be the solution of a set inverse problem here using pagefun with @mldivide might be able to get some speedup if your arrays are large enough. Hope this can help a bit in getting started.

Best,

Kristoffer

I tried to do something like this, but I figure that this depends on the type of problems you're solving. I'm working with a system of several coupled ordinary differential equations, integrated in respect to time, using the integration function ODE15s. This integration would be the most time-consuming step of my algorithm. It runs just fine, less than 5 seconds, if you are just executing the model with defined parameters. But when I wanted to adjust these parameters, using fminsearch for example, this optimization could take hours.

The problem of integration is that the next step depends of the former. So there is no way you can speed-up this computation by using the GPU. Nonetheless, Matlab has several parallel functions such as parfor (see link below). These functions will take advantage of multi-core CPUs (number of workers is equal to the physical number of cores).

So, in theory, one could use parfor to carry out the optimization, with different sets of initial value (each step of the for loop would be a different experiment), in the different cores. In theory, because I didn't try it myself yet, but I don't see why this wouldn't work. You can use the same idea for other time consuming tasks, by using parallel computation with your CPU (not GPU).

https://www.mathworks.com/help/parallel-computing/index.html?s_tid=CRUX_lftnav

I'm posting below a presentation of my project, just to give you the idea (I have a flowchart of my algorithm somewhere in the middle of the presentation): Presentation Modeling of Polymerization of Methyl Methacrylate in Homogen...