From: Axel Kohlmeyer (akohlmey_at_gmail.com)
Date: Wed Nov 07 2012 - 11:05:48 CST
On Wed, Nov 7, 2012 at 5:23 PM, Aron Broom <broomsday_at_gmail.com> wrote:
>
>
> Perhaps AMD graphics cards will promote the kind of competition needed to
>
unlikely. AMD can't affort to keep up with nvidia. they're already
struggling on the server CPU market. i'd expect them to focus
more on niche markets to sidestep intel with its almost infinitely
deep pockets. more likely that intel will create some heat for
nvidia, if they ever get their act together.
> get the better features without the massive costs? It seems like OpenCL
> 1.2 is no longer really awful compared to CUDA, no?
>
OpenCL is less of a problem. the drivers are.
thanks to a donation from AMD, i have been
able to set up a test machine at temple, that
was effectively identical in the CPU/host config
(dual tylersburg chipset, dual quad core westmere CPU,
4 full bandwidth PCI-e 2.x GPUs. the best hardware
config for optimal GPU performance you could get
at the time (pre-PCI-e 3.))
single GPU performance with OpenCL (using LAMMPS)
is pretty good, but as soon as you try to oversubscribe
GPUs or try to access multiple GPUs concurrently,
the comparable nvidia node leaves the AMD node in
the dust. this means, there are too many locks and
other serialization left in the AMD drivers. also the
overall reliability was less. i need some hacks to make
the machine not accidentally hand over the CPUs as
OpenCL device (which the AMD OpenCL drivers support,
but leads to a horrible performance or crashes)
i remember the nvidia drivers being in a similar shape
a few years back and nvidia folks telling me that they
first put all kinds of safeties into the drivers and libraries
and are very slowly pulling the plugs and checking which
API are going to be exposed over time.
so, i'd expect that for AMD via OpenCL the impetus for
running on single GPU desktop-like machines with as
much code ported to the GPU is even higher than for
nvidia hardware and CUDA.
axel.
This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:22:43 CST