Status
Not open for further replies.

Primative-Screwhead

Distinguished
Mar 7, 2011
5
0
18,510
Hello, I was wondering about a crazy idea I had over the weekend. On the EVGA Classified SR-2, would it be possible to use 3 NVIDIA Tesla C2070 CUDA GPUs and 4 Quadro 6000 video cards in 4-way SLI if I get custom water blocks made to reduce them to single slots? One of the things I am unclear on is whether or not the C2070s would have to be in the SLI as well, per NVIDIA 'Personal Supercomputer' specs.
Disregarding practicality issues like power supply, would this even be remotely possible?
 
Solution
^ Same here, I'd be really interested in what this setup is needed for.

I think it is possible provided that the other issues (power, large enough case, cooling, etc) can be taken care of.

If you want to use GPUs for CUDA then you don't want them in SLI, because SLIed GPUs appear to CUDA as a single GPU, preventing you from running computations on the other GPU(s). At least that's what I remember from reading the CUDA documentation.

jprahman

Distinguished
May 17, 2010
775
0
19,060
^ Same here, I'd be really interested in what this setup is needed for.

I think it is possible provided that the other issues (power, large enough case, cooling, etc) can be taken care of.

If you want to use GPUs for CUDA then you don't want them in SLI, because SLIed GPUs appear to CUDA as a single GPU, preventing you from running computations on the other GPU(s). At least that's what I remember from reading the CUDA documentation.

 
Solution

Primative-Screwhead

Distinguished
Mar 7, 2011
5
0
18,510
Thanks for the replies.

As for the reasoning behind it: GPU computing is a very fast and efficient way of doing something that I probably shouldn't admit to doing. :lol:

But in seriousness, I have a lot of specialized applications for my work that are optimized for CUDA, and it would save me a crapload of time if I could have a Tesla PSC at home, but I don't want to sacrifice gaming performance on my free time. I have already figured out the power, case, and cooling, and I expect to be coming into some real cash soon, so don't consider money an object.

Upon further inspection, it appears that the Quadros would not work in quad SLI. They don't seem as though they would accept a 4-way SLI bridge. What about GTX580s? I think I once heard somewhere that GeForce video cards were not recommended for Tesla PSC applications, but I cannot remember where I heard this or whether or not it is reliable information.
 
580s only work in up to 3-way, as far as I know. Also, why would you want quad-SLI? Why not just run a single 580 and then a bunch of CUDA cards? The 580 gives plenty of gaming performance unless you're running at absurd resolutions.

(Needless to say, the PSU requirement for this will be staggering)
 

Primative-Screwhead

Distinguished
Mar 7, 2011
5
0
18,510

I do believe that 580s work. I've seen many 480s in quad SLI. If there is something about the 580 that restricts that, could you tell me what it is?

Absurd resolutions? Two monitors at 2560x1600. I don't know how absurd people would consider that, but it seems pretty taxing to me.

As for the PSU, I'm thinking a Strider 1500 watt coupled with another lower wattage unit (1000 or so, just to be safe).
 

jprahman

Distinguished
May 17, 2010
775
0
19,060
Geforce GPUs are not always the best choice for HPC applications that require double-precision arithmetic because Geforce GPUs have had their double-precision arithmetic performance intentionally reduced by Nvidia. They also do not feature ECC memory that can reduce the possibility of a memory error from corrupting results, although this isn't a big deal unless you are running the cards as part of a mission critical application. So Geforce GPUs are fine for HPC, unless you are doing alot of double precision arithmetic.
 

jprahman

Distinguished
May 17, 2010
775
0
19,060


So you're trying to decrypt classified documents. lol :p

According to Nvidia the GTX 580 only supports 3-way SLI, however, some people have gotten GTX 580s to run in Quad SLI, but it's not guaranteed that it will work and it is likely that there could be bugs. I would just recommend 3-way or even 2-way SLI for the GTX 580s. After you go past 2 cards the scaling gets worse and worse, and by the time you get to 4 cards the law of diminishing returns means that performance just isn't worth the cost and heat generated by 4 GTX 580s.
 

Primative-Screwhead

Distinguished
Mar 7, 2011
5
0
18,510
Thanks again.

I guess I could try 3 580s and 4 Teslas (I'd hate to waste a PCIe x16 slot), unless there's a problem with that. I've never heard of a system with more than three Tesla GPUs, but if they are not in SLI anyway, then I can't see why it would matter how many.

EDIT:
Actually, I might wait for the 590:
http://www.overclockers.com/gtx-590-design-specs-leaked/
It's up for a 4-way.
 
Well, i dont see why you would need to have both the speed and the processing power (mayb im wrong but quadros are not for gaming, but for prossesing).

You might want to forget about that and check Nvidias top (like 16 Gpus in sli?, something like 18k dollars for that one)
 
Status
Not open for further replies.