Building a GPU computer for MATLAB

Our lab would like to build a GPU computer to run a number of in-house codes in MATLAB. The codes model N^2 type problems and require lots of memory to run. We would like suggestions on what sort of machine we could build to get the most out of our budget. What should we take into account when selecting GPUs, motherboards, etc.? Does MATLAB share memory resources across GPUs, or would our codes be limited to the resources on a single card? What are the common hang-ups users come across when running MATLAB codes on GPU systems or in designing such systems? These are the kinds of things we are a bit lost on. Any help would be greatly appreciated!

Answer 1

[Jason Ross]

A few questions:

• What is "a lot of memory"?
• Do you have a license for the Parallel Computing Toolbox?

A few things I can tell you:

• The memory resources are not shared.
• If you build a system that is going to actively use many GPUs continuously, you are quickly going to want to move the system to a server room or other unoccupied space to avoid the noise of the many fans running.
• Consider a rack mounted system if this is going in a server room, or at least verify that the system you select has the ability to fit in a standard 19" rack. Many workstations have rack adapters available.
• GPUs consume a good deal of power (200W or so), so the machine you spec out needs to have a good power supply in it.

There are a couple of web pages from MathWorks and nVidia that may help you select systems and answer your other questions:

http://www.nvidia.com/object/tesla-matlab-accelerations.html

• This has a couple of system configurations that might provide a good starting point for you, and what GPUs to spec out.

http://www.mathworks.com/discovery/matlab-gpu.html

• Links to documentation, examples and demos.

[Walter Roberson]

I also recommend that you have a look at liquidnitrogenoverclocking.com as they make high-performance systems and specifically test MATLAB. Whether you buy from them or not, their site is a useful resource for understanding the considerations for high performance computing on this class of machines.

Answer 2

[Tommy Sebastian]

The saved output files are about 1.5GB and MATLAB runs out of memory when I run the code on anything other than (1) a 64-bit machine or (2) a 32-bit Windows XP machine with the /3GB flag set in the boot.ini file. So I'd say between 2-3GB of memory. Also, we do have the PCT.

[Jason Ross]

It seems that you have a couple things going on.

1. You want to be able to solve your problem and need enough RAM. That would imply that the 64-bit machine with a decent amount of RAM is what you need.

2. The second thing is that you want to increase computational speed using the GPU. A GPU will have a certain amount of on-board RAM (4-6 GB is not uncommon), but you need to move the data over there, do something with it, and move it back. I'd take a look at the supported functions in the documentation and make sure that they can perform the operations you are looking for.

You might want to also look at getting a GPU and putting it in your existing 64-bit machine to get an idea of if it's going to work for your problem.