Multi-GPU computing in MATLAB

I'm considering getting either:

1) one GTX 690 (4GB of RAM shared between two 1536 core GPUs, which share the same PCIe bus)

2) two GTX 680s (2GB of RAM on each 1536 core GPU, but each has its own PCIe bus).

3) one GTX 680 (4GB of RAM, 1536 cores, one PCIe bus)

Each of these options has 4GB of RAM, but my worry is that with option 2, if X is a 3GB array:

X=gpuArray(X)

will not work because I don't think MATLAB supports distributing an array over two separate GPUs (how long do we have to wait for MATLAB to support this type of thing ?)

This also makes me worried about option 1: would MATLAB let me put the 3GB array on the GTX 690's 4GB of RAM that's shared between two GPUs ? If I did have a 3GB array on that shared RAM, would MATLAB be able to perform BSXFUN on it using the 3072 cores that are spread between the two 1536 core GPUs in the GTX 690 ? Or would it only know how to use one of the two GPUs to perform calculations on this array ? Has anyone tried to put a 3GB array on a GTX 690 ?

The use named 'kyle' said "it will be possible, in a future release, to use multiple GPUs to solve one large problem" more than a year ago on this forum thread: <http://www.culatools.com/forums/viewtopic.php?f=15&t=873>

How close are we to seeing the realization of that statement in MATLAB ?

If one googles "multi-GPU matlab" they will see that I am one of MANY people interested in answers to the above questions ! =)

[Amit Hmath]

Hello All,

Please Checkout our Novel way of GPU computing at: https://www.vivid-sparks.com/

Many Thanks!

Answer 1

[Edric Ellis]

You're right that each gpuArray can reside only on a single GPU, and that a single MATLAB instance can access only a single GPU device at a time. However, you can use matlabpool to launch additional MATLAB workers and have those each use a different GPU device.