Hi
LAMMPS will likely be able to run on MIC processors as is, but for optimal performance we expect to need a specific package along the lines of the GPU, USER-CUDA and USER-OMP package. In particular one can expect that to be necessary (as Axel suggested) if your system doesn't fit on one MIC card, and you need communication between multiple of them.
best regards
Christian
-------- Original-Nachricht --------
Hi
LAMMPS will likely be able to run on MIC processors as is, but for optimal performance we expect to need a specific package along the lines of the GPU, USER-CUDA and USER-OMP package. In particular one can expect that to be necessary (as Axel suggested) if your system doesn't fit on one MIC card, and you need communication between multiple of them.
there also is the issue of having to write force
kernels with proper SIMD/AVX support for
maximum performance.
at least from my limited experiments and
based on the "theory" i learned about writing
such code, the current data structures
in LAMMPS don't lend themselves very
well for efficient vectorization via such SIMD
instructions due to lack of alignment
and requiring gather/scatter operations.
axel.