Nvidia GPU card for parallel computing, personal use

I posted this earlier being somewhat naive. Let me rephrase:

Does anyone know the best Nvidia card for scientific parallel computing, other than the Tesla?

I am looking to buy a workstation, hopefully near my $2,000 budget, for scientific parallel computing. I am aware the Tesla is tailored for scientific computing, however the card alone is more than my budget.

Any help with this would be most welcome. I have looked at some GeForce GT/GTX cards but am unaware if these will suffice for my needs. To offer some guide as to the computing I will be doing, the math is essentially linear systems (to solve the a system discretized by the finite element method).

Given that the GeForce 500 series can have some (unpredictable) compute errors (see GPU in state where results are not reproducible), and the new GTX 680 is more oriented to gaming than to compute (although a 30%-60% speedup compared to a GTX 580 is reported), I would try to get my hands on a GTX 480 (or maybe a 470). In case you need to transfer a lot of data from main (CPU) memory to GPU (relative to the amount of computations you expect to do on the GPU), I would go for the GTX 680, since it is the only GPU that supports PCIe 3.0, which (maximum theoretical) bandwidth is twice as large as PCIe 2.0. Make sure you have a motherboard that supports PCIe 3.0 in this case.

A nice Wikipedia page to compare NVIDIA GPUs: Comparison_of_Nvidia_graphics_processing_units.

From those posts the problem came with the over clocked cards. It seems that the memory errors disappeared when the memory was down-clocked or the voltage was increased. Those errors do not affect games so much, but they change the results. Somebody on the forum reported as well a problem with an overclocked 580 GTX. It was working on windows but not on linux.

Yeah, I don’t know much about these GTX 580 problems, but if that turns out to be an issue with overclocking, then I would highly recommend a standard clocked GTX 580. The Tesla cards add features that you probably won’t need if you are just starting out with CUDA. Tesla C2050 cards have 4x faster double precision throughput (ignoring clock rate differences), and are tested for 24/7 compute usage. However, in other things, GeForce and Tesla are equivalent, so I would not spend the extra money until you learn more about CUDA and can evaluate if it is required for your use cases. Definitely don’t bother buying a C1060. The GTX 580 is better in so many other ways than the old Teslas.

Thanks a lot, that was a HUGE help. I was able to configure a full workstation on Velocity Micro with quad cores and a GeForce GT 520 for $1300. It was only $80 to add the GT 520, but for the GTX 580 it’s $675 which is pretty high. On the wikipedia comparison, the GTX 580 is obviously the better card but I don’t see the specs being $600 better. Thoughts?

All of the GeForce add-ons were in the 500-series, I guess they are phasing out the 400’s because I can’t seem to find them anywhere.

I’d recommend not going with any Geforce GT card. It will be a huge disappointment when you discover that the GPU cannot even get close to CPU speed. And particularly in the case of the GT 520 with only a single SM (streaming multiprocessor) it will also guide you in the wrong direction regarding performance metrics - you will probably have to completely rewrite your code to get to decent speed on better cards later.

The GTX 560 Ti 448 core edition is a good choice for CUDA.

Tera, wow, thank you for the recommendation. GTX 560 Ti 448 is only $270 on amazon. I’ll probably end up going with a custom build with a price like that-- I haven’t seen any workstations thusfar that have the option to add the 560 Ti 448. I’ve never done a custom build before. My area is software, not hardware. This should be interesting.

Usually my colleagues and I just go the the computer hardware shop nearby and buy a regular (gaming) PC and use it as a compute-server. It is much cheaper than a custom build workstation, and if it breaks it was cheap enough to replace with the latest and newest hardware. Just tell the guys at the computer store that you want a PC with a high-end Intel CPU, a lot of memory (8GB or more, it is cheap anyway) and 1 or 2 GPUs that all run at maximum PCIe speed (important for motherboard) and the computer-store guys can do the math for you.

The issue with the GTX 580 and GTX 570 cards is NOT only with overclocked cards. We underclocked our 16 GTX 570 cards, and at least one card will show an error every day. GTX 560 might not have the same problems, since it has a GF114 chip instead of a GT110 chip for the GTX 570 and GTX 580 (but I never tested a GTX 560 for errors, so don’t blame me ;) ).

Try to increase the voltage on the memory individual, or decrease the memory speed only.