Multiple Devices with GTX 295

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

On the GTX 295 card, there are 2 separate GPUs on it, each with 240 cores. I was able to get code working using the default setting without calling cudaSetdevice(), which uses GPU #0. However, I try to run the code after adding cudaSetDevice(1) to use GPU #1, I get a significantly different numerical result. I am currently using GPU #1 as my display device in X11 as well. There are known issues with using CUDA with a display device, but I’ve only seen performance issues with slower computation. Could it also affect its numerical precision and accuracy? Could there something wrong with the card?

2

I use both GPUs of GT295 getting the same results. You can use also together using two host threads.

There are known issues with using CUDA with a display device

Yes, there’s a watchdog that kills your kernel if it takes more than 5 seconds. So kill your X server if using big kernels.

Cheers,

luca