Nvidia for GPGPU but ATI cheaper and better...

B

blisk

Limp Gawd
Joined
Feb 4, 2005
Messages
183
I would really like to get into GPGPU programming and I know matlab has CUDA acceleration which has me leaning strongly towards an Nvidia card. However their cards really do seem to have some issues with heat, and fan noise on the 560ti and 570 cards I'm looking at. The other thing is that there are 2GB versions of the 560ti and 560 however only one card way above my price point with 2.5 GB for the 570 ($400). If it was a next gen card I'd be aright with spending $400.. but this gen is definitely winding down and it seems like Nvidia and ATI are basically getting rid of inventory at a pretty profitable price for the holiday season while the new lower yield (initially) 28nm parts somehow works out that it'll be coming out in January. 28nm scaling issues seemed to work out pretty well for them.

I haven't seen benchmarks but for AA at 1080p it seems that video games do suffer by not having a lot of ram... but I can't find any benchmarks showing significant differences in BF3 for 1GB vs 2GB for Nvidia. And so I'm strongly considering just getting a 6950 2GB that i can unlock to a 6970.

The other option is to wait but there are really very few hints of when the next gen cards are coming out and their availability. I would like to ideally have a card no later than a month from now in my PC so I can learn a bit of the programming over break. Ideally it'd be great to wait for these next gen cards... but I don't think the timing will work out.

Ideas? Maybe I should pick up a cheap/used CUDA Nvidia card that I can use as a learning tool?
 
mannyman said:
There is a guy selling a 4870 for 55 shipped, I imagine that will have enough horse power to tide you over.
http://hardforum.com/showthread.php?t=1648782&highlight=4870
Click to expand...

Naw I have a 4870 currently... its ok.. but i upgraded everything else but kept the video card the same. Its pretty sure though that its holding me back. And OpenCL isn't easy to learn nor used that much at this moment for GPGPU programming it seems which is why I'm leaning towards an old nvidia card to program on.
 
1GB cards do just fine at 1080p. I think getting a dirt cheap CUDA card separate from your gaming card is a good idea though. You might decide GPGPU isn't for you. The only problem is that low end cards are probably slower than your CPU due to the additional pcie transfer latency and the performance might turn you off from CUDA/OpenCL programming.
 
The Nvidia's being better depends a lot on the type of code.
The simple opposite example that everyone knows is//was Bitcoins.
 
Off course the cards will run hot. Consumer cards are made for speed not precision like the workstation cards which is lower clocked. Get a geforce card and just down clock it to run cooler like a quadro card.
Nvidia cards run better because nvidia have the advantage with the shader performance where Amd got the advantage with the ram connection. But overall the faster shaders gives them the edge
 
You must log in or register to reply here.
Back
Top