[?] Best low profile CUDA GPU, double slot ok,

[allenpan]

[?] Best low profile CUDA GPU, double slot ok,

i need it for my other SFF CUDA runtime, too bad they dont have low profile Tesla, else i will be all over that..

prefer single slot, but double slot is ok too...

all my codes are HW 2.0,and CU4.2 so anything in fermi or higher will work, but i never really test code on kepler, wondering if is worst or better, as many poeple claim Nvidia on purposely cripple geforce's Kelper for CUDA computation

 

[BallaTheFeared]

Are you looking for a workstation card with high SP/DP performance, or just looking to access CUDA?

 

[allenpan]

either is fine, i do have a high end tesla for dev, but i need s low profiel one for runtime

 

[BallaTheFeared]

Something like a GT 430?

 

[allenpan]

too weak, that is why i am asking the "best", i can find something like GTS 450 low profile from Sparkle and Pailt but cant seems to buy them anywhere, but i rather to have even more cores,

my multi-targeting system code's scalability:
Intel C2Q 9200 = 4 fps
Geforce 520 = 5fps
Geforce 430 = 8fps
Geforce 560Ti (336) = 27fps
Geforce 570 = 32fps

so i assuming 450 with 192 cores scale about double of 430 = 15 fps

now the major questiong is if Kepler will scale better? i see other colleague from other insititue tried 680GTX and saying it is slower than 580GTX, but i dont know if is DP or SP

i only need SP

 

[BallaTheFeared]

The 580 is faster at DP than the 680, the 680 though should be slightly faster than the 580 at SP.

I think your problem is the fact that low profile is going to be hard to find.

I can't even find GTS 450's on ebay.

 

[allenpan]

humm, do u know if there is any performance difference between tesla vs geforce in SP? assuming same amount of cores/clock and memory size is not bottleneck?

also

whats the fastest SP CUDA card out there? i think the worst comes to worst i just leave extend the PCI-e slot and use full size card

 

[dave_the_nerd]

Something like this:

http://www.palit.biz/palit/vgapro.php?id=1895

It's low-profile (or low-profile ready - the card is, but dunno if it comes with the right bracket.)

384 cores is nice. Memory bandwidth with GDDR3 isn't great.

 

[BallaTheFeared]

^ that is basically the Kepler version of the 450. 384/2 = 192

How it performs though I have no idea, look for reviews, typically professional cards are going to be faster because they probably won't be cut as much but nvidia can do anything they want and you won't know until you see an actual review.

 

[allenpan]

i find it very difficult to see any review on tesla vs geforce of sp benchmark

 

[ViRGE]

humm, do u know if there is any performance difference between tesla vs geforce in SP? assuming same amount of cores/clock and memory size is not bottleneck?

GeForce SP performance is not restricted. So a GeForce card shouldn't be any worse than a Tesla card for your needs.

Just keep in mind that NVIDIA doesn't even make low-profile Tesla cards, so this is all theory crafting. If you need a low-profile card, you're basically stuck with GT 640 unless you can find a second-hand GTS 450.

i find it very difficult to see any review on tesla vs geforce of sp benchmark

Instead of looking for Tesla results, look for Quadro results. Quadro cards have the same unrestricted compute performance of Tesla, and it will be much easier to find Quadro/GeForce comparisons than Tesla/GeForce comparisons. The latter is very rare since Tesla buyers aren't even considering GeForce cards.

In the meantime there are a few benchmark programs that have Tesla results, such as CLBenchmark.

http://www.clbenchmark.com/device-info.jsp?config=13067078
http://www.clbenchmark.com/device-info.jsp?config=11975142
http://www.clbenchmark.com/device-info.jsp?config=11905948

Finally, it's worth noting that most of the "Kepler's compute performance sucks" comparisons are against the high-end Fermi GPU (GF100/GF110). The rest of the Fermi GeForce lineup used a significantly different design (Fermi-Lite) that didn't perform nearly as well for compute. So just because a GTX 570 is good at compute relative to a GTX 650 doesn't mean a GTS 450 will be relative to a GT 640/650.