Best Bang for Buck CUDA card looking for the right card for me, on a budget

Currently I am running the following for my system.

CPU: AMD Phenom 9950 overclocked to 3.0GHz
RAM: 8GB DDR2 800MHz
Video Card: Geforce 9600GT 2GB

I am currently doing FDTD work with large matrices in MATLAB and would like to exploit the GPU to speed things up. I’m on a tight budget so have been looking into the gtx 550 and 560 cards.

Can someone make a solid recommendation on what card would work best for me. I was hoping the 600 series would be something I should look at but the review on Tom’s Hardware said that the compute capabilities are being scaled back in the gaming graphics cards. Is this really true? and does that mean I should look at older generation cards like the 400 and 500 series?

Also I am trying to stay below $300, but if the additional performance is there I am willing to go higher.

Thanks in advance

Go for the GTX 560 Ti 448 core edition, it should still be within your price range and is the best CUDA card there.

That was a good thread.

I am using very large matrices so I want a card with a good size of memory on board. So I was thinking something with 2GB. This leaves the 550Ti, the 560 and the 560Ti.

The difference between the 550 and the 560s is the memory bandwidth. How large of a roll does that specification play in how speedy the computations are, or maybe it doesn’t matter at all?

Also if I went for the GTX 560 Ti 448 with its 1280mb of memory and found that the memory size isn’t enough can MATLAB and CUDA recognize cards in SLI and use both banks of memory?

I’m seriously considering a second hand GTX580, either that or wait for GK100 (or whatever is going to be called).

This makes a difference depending on your ratio of arithmetic operations to memory reads and writes. Many CUDA algorithms are memory bandwidth bound, so you should figure out if your intended application clearly bandwidth bound.

That said, without knowing anything about your specific application, I’ve never regretted paying for more memory bandwidth. :)

SLI does not have any impact on CUDA. With two cards in SLI mode, you will still see two separate CUDA devices with their own memory space, and memory transfers between devices will still use the PCI-Express bus. You will have to copy data between the devices manually, or possibly look into unified virtual addressing to streamline memory copies between devices. (UVA might only work for GeForce cards in Linux, I can’t recall at the moment…)