Hi all,
can the cufft lib use multiple GPU’s ? If so, how should that be done ? Will it just run faster automatically, or should I run several simultaniously, or … ?
Is there any documentation about this ?
tia,
wateenellende
Each plan uses only one GPU, you can run different plans in different contexts (threads) to use several GPUs for several FFTs
FFTs is very memory intensive with very little compute. I didn’t always get performance improvement even using concurrent copy and execute with it as running time was much shorter that PCIe communication time, let alone multiple GPUs, so running it on multiple GPUs is going to take you even more time