[gmx-users] Why we are not using GPU to solve FFT?

[Mark Abraham]

On Fri, Aug 22, 2014 at 9:10 AM, Theodore Si <sjyzhxw at gmail.com> wrote:

> Hi,
>
> I wonder why we are using cpu instead of gpu to solve FFT?


While people continue to attach GPUs to nodes with tasty x86 cores, we'd
like to use them. Adding more work to the GPU while leaving the CPU idle
does not improve throughput.

Is it possible to use gpu fft library, say cuFFT to make the FFT used in
> PME faster?
>

Perhaps, on a single node, if someone were to write the code to do it. But
on mutliple nodes on current-generation hardware, a 3DFFT that has to do
data transfer, kernel launch and then go back across PCI to get to
Infiniband to do the all-to-all would be horrendous. The latency of
all-to-all communication already limits scaling without, say, doubling the
length of that latency.

Mark


> BR,
> Theo
> --
> Gromacs Users mailing list
>
> * Please search the archive at http://www.gromacs.org/
> Support/Mailing_Lists/GMX-Users_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
>