Multiple GPUs

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I started to work with 2 GPUs, and to investigate how it works I used the simpleMultiGPU program which is published by NVIDIA and exists in: \NVIDIA Corporation\NVIDIA CUDA SDK\projects folder. Using both GPUs the processing time for this project in my system is 161ms and CPU’s processing time is reported as 227ms. To realize the computing speed of 2 GPUs, I changed the MAX_GPU_COUNT to 1, which originally was 8. By this change only DEVICE 0 will work, and reasonably the processing time should increase 2 times, right? However, the processing time is not changing too much. It is around 168ms!!
Could anybody explain why the processing time is not around 320ms?
Could you please try this program on your system and let me know how much the processing time is for single and multiple GPUs?

Thanks…

2

simpleMultiGPU is a basic example to show you how to write an app for multiple GPUs and will not show a performance boost from having more than one GPU.

3

Thanks for your reply.
Could you please let me know why it doesn’t show this performance? in other words, what is the application in which using of multiple GPUs has a good performance?
Another question…if CUBLAS library can work with multiple GPUs?