3
$\begingroup$

Are consumer grade GPUs (like the NVidia Geforce series) used for solving sparse linear equations systems in a professional setting? For example, by engineers performing finite element analysis?

After all, at least for iterative methods, memory bandwidth is the bottleneck, and a Geforce 780 Ti provides a lot more of this than a Tesla C2075, while also costing a lot less.

|cite|improve this question
$\endgroup$
  • 2
    $\begingroup$ Look at my answer to a related question here. There are other features you get with the Tesla series that may or may not be important for your application. $\endgroup$ – Godric Seer Feb 9 '14 at 18:30
3
$\begingroup$

In a "professional" setting usually you run commercial software only on "certified" hardware, i.e. hardware on which the vendor has tested his software and is willing to provide user support. "Certified and supported" GPU's lists are usually very short...: there is really no technical reason, just commercial ones.

In my experience, high end FEA solution vendors charge such horrible high sums for licenses that clients do not care for the cost of the GPU cards and just buy the supported ones.

This said, if you are going to use "consumer grade" GPUs, you should test if, under sustained load, accuracy of the solution and performance is preserved.

|cite|improve this answer
$\endgroup$
  • 1
    $\begingroup$ Please see Godric Seer's answer on a different question for a discussion of the importance of ECC with respect to solution accuracy. $\endgroup$ – Stefano M Feb 9 '14 at 18:45
  • $\begingroup$ I was just about to say the exact same thing. This is especially clutch in problems where the condition number (ratio of the max to min eigenvector) of the sparse matrix is is high. $\endgroup$ – meawoppl Feb 13 '14 at 5:04
1
$\begingroup$

Sparse matrix multiplication (that is, sparse matrix inversion/solution) is a huge area of interest/research in the GPU community. Currently things are looking pretty good, but it's not anywhere near where it is in some other fields. (This is really hard.)

At the nuts-and-bolts level of an analyst using a practical FEA code, it's not there. All the big players are waiting to add/push support until the situation is a bit clearer.

|cite|improve this answer
$\endgroup$
  • $\begingroup$ I'm not sure I understand your second paragraph. Are you saying that many FEA software manufacturers don't support GPUs yet? What would make the situation clearer? $\endgroup$ – Adrian Willenbücher Feb 13 '14 at 7:57
  • $\begingroup$ Yes, I mean none of the popular vendors support GPU in any meaningful way. The situation will be clearer when we find a way to write GPU code that isn't as dependent on the specific hardware you're running it on as it is today. $\endgroup$ – Mike Feb 13 '14 at 15:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.