Hi, On Wed, Sep 2, 2015 at 9:34 PM Fabricio Cannini <fcann...@gmail.com> wrote:
> On 27-08-2015 10:03, Szilárd Páll wrote: > > A few more things to add: > > - we don't have CPU SIMD kernels for Power8, so comparing the plain C > > kernels against AVX kernels on AMD is not exactly fair; > > That was our hypothesis after seeing the results. > > One of our goals was to assess the readiness of some softwares that are > popular among our clients; See if is it too early for them to consider > purchasing POWER8 hardware. > As Szilard notes, we have not yet optimized POWER8 for CPU-only runs. It will likely be optimized by the 2016 version of GROMACS. POWER8+CUDA is OK now, and will improve further. Mark > - as Mark pointed out, to use resources efficiently, you'll have to manage > > thread affinity on Power8 (by default you'll need a stride of 8 to use > one > > thread per core, a stride of 4 to use two, etc.); > > - for a realistic performance measurement use ~100-1000 atoms/core; the > > "0000.96" benchmark in a 960 atom input, running that cross 192 threads > > makes little sense. > > Are any of the inputs in the 'water' benchmark useful for this purpose? > > > What motivates your choice of compiler flags? > > We had reports of good performance using these flags in a POWER7 machine. > > > > TIA, > Fabricio > -- > 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-requ...@gromacs.org. -- 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-requ...@gromacs.org.
