Sign-in / Sign-up
Your question

CUDA card, NVIDIA GT 640 ddr5 same gpu chip different bandwidth for Asus and gainward

Tags:
  • Bandwidth
  • Gainward
  • Graphics Cards
  • Graphics
  • Nvidia
Last response: in Graphics Cards
February 6, 2014 2:21:16 AM

Hi everybody,

i would like to start doing CUDA programming.
And i was thinking about graphics card NVIDIA GT 640,
because it is the only affordable card, that has compute capability of 3.5.

There are different companies selling video cards with this gpu-chip
the Gainward one, is listed as x8 2.0 but the Asus and Gigabite ones are x16 3.0.

The system has x16 2.0. they all should be ok.

i was just curious about the difference in performance due to the different bandwidth x8 vs x16? From other threads i understood that for games, and stuff there should not be to much difference, but how is that for cuda programmed linear algebra stuff. Any one some experience about how much of the available bandwidth that is using?

Thanks in advance for any helpfull commends.

More about : cuda card nvidia 640 ddr5 gpu chip bandwidth asus gainward

a c 79 U Graphics card
a c 76 Î Nvidia
February 6, 2014 5:13:00 AM

The Gainward one you're looking up looks like an earlier revision GT640. Gainward definitely sells a PCIE3.0 x16 640.

I would say they all plug and play the same, but I'd get the one that supports the higher PCIE bandwidth just to be safe, as they'll run at x16 in PCIE2.0.
m
0
l
a c 84 U Graphics card
a b Î Nvidia
February 6, 2014 5:17:59 AM

You can safely get the Gainward if you want.
I have the Gainward 650, and it's 3.0x16 GDDR5.
Just that it does not come with any accessories.
m
0
l
March 14, 2014 9:58:13 AM

Thanks for your answers.
later i found out that the compute capabiliy 3.5 wich is really rear among the nvidia cards, doesn't really say anything about the speed, just about the supported technology.
So the particular 640 of interest allows to test the cuda 3.5 stuff, but unfortionatly very slow.
So i guess ti would be like a wheelchair having the right specs to fit a F1 tire. yes you can test if it works, but not really anything interesting about speed up. Because i wanted dynamic paralellism (only cuda 3.5 and up) i guess i have to save some more, and be very lucky on ebay, geting a used titan.
m
0
l