Quick 9800 GX2 question regarding SLI memory

Accelerated Computing CUDA CUDA Programming and Performance
1

Firstly, am I right in thinking that the GPUs in the 9800 GX2 have compute capability 1.1?

And secondly, a question about 9800GX2 memory usage: Since it’s two cards connected via SLI, is it possible in cuda to use only 1 GPU, and access the full 1024Mb of memory for the computation?

I need to know because it is quite possible that some of our computations will require more that 512Mb, and so we’d like to use all 1024Mb from one of the GPUs for a single computation. Is this possible?

Thanks

nodlams

2

9800 GX2 is the G92 chip, so yes it is compute 1.1. No, you cannot access all 1024MiB from one card. CUDA sees each GPU as an independent CUDA device with its own memory bank. SLI is just used to transfer rendered pixels from one card to the other so it wouldn’t be useful in your use-case even if it was available for use in CUDA. For more than 1024 MiB on a card, you need Tesla (or maybe Quadro? I’m not sure).

3

Looks like the Quadro FX 5600 is a G80 with 1.5 GB of memory. Costs a lot more than Telsa, though, so it would only make sense if visualization were important to the problem.