Hi Erik, Thanks for the quick reply! It certainly sounds like Epyc may be an interesting and competitive option.
Best, Geraint On Mon, 20 Jan 2020 at 18:19, Erik Schnetter <[email protected]> wrote: > Geraint > > Regarding BLAS: I recommend using OpenBLAS instead of MKL there (but > haven't compared performance). There is a thorn > ExternalLibraries/OpenBLAS (part of the ET, but not enabled by > default) that interfaces to OpenBLAS and/or builds OpenBLAS if it is > not already available on the system. > > I am not aware of any issues. Regarding benchmarking, thorn > CactusUtils/Vectors supports SIMD vectorization, and it should already > support Epyc. From the technical discussions on the web it seems that > they should be well suited for the ET. In particular the large L1 > instruction cache should be quite beneficial. > > Thank you for the pointer to the Prace best Epyc practices. > > -erik > > On Mon, Jan 20, 2020 at 11:50 AM Geraint Pratten > <[email protected]> wrote: > > > > Hi everyone, > > > > Does anyone have any experience running on AMD Epyc cores? It seems that > these are becoming quite trendy and I found a pretty decent best practice > guide that Prace released > > > > http://www.prace-ri.eu/best-practice-guide-amd-epyc > > > > However, whenever I've compiled ET recently its typically been on Intel > cores + Intel compiler + Intel MKL etc. However, Intel MKL is notorious for > being subotpimal on non-Intel cores. Does anyone know of any possible > issues that could arise when running ET on AMD Epyc cores? Similarly (and > likely optimistically), has anyone had the chance to benchmark the > performance of ET on the AMD cores? > > > > Thanks! > > Geraint > > _______________________________________________ > > Users mailing list > > [email protected] > > http://lists.einsteintoolkit.org/mailman/listinfo/users > > > > -- > Erik Schnetter <[email protected]> > http://www.perimeterinstitute.ca/personal/eschnetter/ >
_______________________________________________ Users mailing list [email protected] http://lists.einsteintoolkit.org/mailman/listinfo/users
