Hello!

I always thought that DP performance was the measure of how fast a GPU solves scientific problems, as that is the number of calculations it can do per second. However, after comparing the performance of many systems from profiles on Boinc and here on the grid, I am a confused.

Specifically, I am unsure what the purpose of the ultra expensive nvidia Tesla cards might be, as they perform like a GTX 780 Ti, even though they should be many times faster. The same is true for other high priced cards like the Titans.

For example, look at these 3 profiles:
http://www.gpugrid.net/hosts_user.php?userid=64884
http://www.gpugrid.net/hosts_user.php?userid=96729
http://www.gpugrid.net/hosts_user.php?userid=71287

Here we see Titans, GTX 780 Ti, and Tesla K20C, but individual WUs are solved faster with the GTX. The GTX has a DP performance of 210 GFlops, while the Titan and Tesla are supposed to have over 5 times that amount. Previously I assumed they'd solve a WU at least 5 times faster. Shouldn't they, especially given their price and purpose as supercomputer cards?

Any insight into this would be appreciated. I plan to build a GPU farm for boinc/GPUgrid next year and have considered to go for Teslas or Titans, as their power consumption is so low per DP GFlop. I reasoned that I would require few cards to get the same amount of GFlops that I would get for many more GTXs, thus saving space, noise, additional hardware and electricity costs over the course of the farm's lifetime.

Thanks!

DP is not needed here, so you don't have to look specific for that, unless you want to do MilkyWay at home. Even between WU's here there is a difference in speed, so I guess is partly due to programming as well. And the set up of the PC, the CPU, has little do with it too.

These Tesla's are so expensive as they are made as a "special". These cards have different drivers and can do some special tricks with the software. As an example: Tesla in a PC for MRI or CT scans can zoom in on an area the radiologist or the medical doctor is interested in, and then in another and better resolution then the rest of the picture. So you see a picture in a picture with a different resolution, a higher and thus sharper picture. That window can be drags everywhere in the screen and will stay sharp. You need a bit decent monitor for that as well.

And perhaps there are other reasons why these cards are so way expensive, but others have to tell you why.

Titans and Titan-Z have more memory and area bit of a marketing stint. But they had a huge price drop in the Netherlands.

What I see is that a GTX780Ti is still the fastest, or nearly the fastest. But are hard to find already.
____________
Greetings from TJ

Thanks TJ, that explains it! The SP performance is mostly the same between these cards an correlates correctly to the results I'm seeing. I had not considered these projects all just use SP :)

Actually this project does the DP needed on the CPU.
Other projects like Milkyway and PrimeGrid do use the GPU for DP, so Titans and Teslas perform much better there than GTXes.
If you intend to build a crunching farm for GPUGrid, the best choice is the GTX970 and the GTX980, as these are almost as fast as a GTX780Ti, but consuming only the 2/3rd electricity.
The other factor for GPU performance is the Operating System, as Windows XP and Linux is faster by 5-10-15% (depending on the WU batch) than Windows Vista, 7, 8, 8.1, 10 etc.

Thanks for the advice, Retvari, I'll consider going with the 9th generation, then. Although there will be a 980 Ti next year (in Feb, I think), so I will be waiting to see their final specs and price.
I suppose the new Tesla K80 with its 2.91 DP TFlops would perform like 13-14 GTX 780TIs on those DP projects, while using the power of a single card.

@Apes: I'm techy, but I've never built a farm/cluster, so it will be a big learning experience for me. I've been reading up on tutorials: Luckily, the bitcoin craze has produced a lot of tutorials for home made clusters that I can dive into.
Sometime next year, probably in the second half, I will start with 2 old systems to learn how to turn them into a cluster correctly, then I can go bigger. Eventually, I want to end up with a full server rack, if I can fit each unit well into one. Money doesn't play a role for me, which makes that part no concern.