Hi,
I read the thread in the FAQ for hardware selection from user Richti and I'm wondering if it is still up-to-date, that 'DELL T7600 2x8' is one of the best machines for LIGGGHTS simulation. It would be nice if there exists something like an ASIC just for the DEM calculations. If some of you have some information about best hardware for LIGGGHTS I would really be interested in.
kind regards
Joshua
richti83 | Tue, 07/14/2015 - 15:48
Well, this hardware is a
Well, this hardware is a little bit outdated (I think we own this machine over 2 years now).
If I would have the money, I would buy a Dell Precison Rack 7910 with 2x Intel Xeon E5-2699 v3 Prozessor (2,3GHz, 18C, 45MB Cache, 145W, Turbo, HT) to have 36 cores in one "node" / workstation.
I think it's not suitable to produce an ASIC only for DEM because the use case is too manifold. I see two opportunities for a speed up in far future: Intel Xeon Phi co-processors and GPGPU acceleration. Both are supported by native lammps but there is no manpower to extend the granular code base to be used with hardware acceleration. And because both known accelerators use very low clock frequency it's not clear if there would be really a speed-up for poly-disperse particle systems.
Richard Berger did a big step by releasing an openMP multithreading approach [1] which seems to be 20% faster than the known MPI based implementation, but it's a lot of work to maintain this branch and I think without a bigger community there is no way to exceed current limitations in computing "costs" ...
But hey, LIGGGHTS is the fastest DEM code out there as shown in many benchmarks.
[1] http://www.cfdem.com/comment/15531#comment-15531