GeForce GTX 295 vs. 285 for CUDA development

I am currently using a GeForce 285 for CUDA development but I am considering an upgrade to the 295. I wanted to ask you guys what are the pros and cons of the 295 vs. the 285.

I understand that the 295 is basically two 285s stuck together. Does this raise any problems? Are the cards wider, do they require more power, and are there any additional programming issues? Do the two GPUs show up as one CUDA device or two?

Thanks.

Yes, the 295 requires more power than the 285 (because it is two GPU’s instead of one).

The 295 shows up as two separate devices in CUDA. If you want to take advantage of both devices, you’ll need to implement some special programming techniques (see the multi-GPU example in the SDK).

It’s actually more like two underclocked GTX 275s stuck together. Each device has a 448 bit memory bus, while the GTX 285 has a 512-bit memory bus. For a program that can easily use multiple GPUs, the GTX 295 will do better. However, if you can only use one CUDA device, the GTX 285 will be much better.

As the GTX 295 has 2 GPUs on a single card, is the bandwidth provided by the PCI-e bus splitted in two ? (meaning you have half bandwidth for each GPU)

Larry

The PCI-e bus is switched, so each GPU can get up to the full available PCI-e bandwidth, but in the worst case situation with both GPUs simultaneously trying to consume as much bandwidth as they can, the effective bandwidth per GPU will drop to a bit less than half.