Matlab with a Titan Z

How does Matlab handle the dual-GPU Titan Z? Can you target each individual GPU using something like gpuDevice(1) and gpuDevice(2) (and is it possible to do so in different Matlab sessions)? If not, is it possible to see any performance increase over a single Titan Black (normalizing for clockspeed differences) using a Titan Z as a single gpuDevice?

For specificity, consider the task of repetitively multiplying/adding several large matrices together, eg. something similar to:

% Silly Example Code
for i=1:2000
    matrixA = matrixC*matrix1;
    matrixB = matrixD*matrix1;
    matrix1 = matrix1 + matrixA.*matrixB;
end

Where you might want to do this over multiple completely independent runs (using separate Matlab sessions and compute devices). If a Titan Z was faster than a Titan Black on an individual run or if it could handle two independent runs simultaneously it would be useful. It would only be problematic if the Titan Z wasn't faster than a Titan Black in any respect.

Answer 1

[Joss Knight]

The spec would seem to indicate that each half of a Titan Z is slower than the Titan Black, so it does depend on how you use it as to whether it can be faster.

To use your dual GPU in parallel you can indeed run two separate MATLAB sessions specifying gpuDevice(1) and gpuDevice(2). But the proper multi-GPU way is to use a parallel pool with 2 workers. Get started with this blog post.

[SimpleYeti]

Yeah, the Titan Z is slower (but uses less power) because it's clocked lower.

As far as I can tell explicit multi-GPU would require a lot of work, as it'd go from a single statement matrix multiply to developing an algorithm to efficiently divide chunks among GPUs. It's much simpler to just run separate jobs simultaneously, which will work so long as both GPUs on the Titan Z can be addressed separately.

Thanks!