Dear all,
At my lab we are thinking of buying a new cluster (not too large) which would combine GPUs with either processors or Xeon Phi. However, some questions need to be answered before we decide if we buy procs or the MIC.
So far I have been using LAMMPS with GPUs (K80) and I must say it works well. I observed in general (at least for my problems) speed up factors of about ~ x5 in comparison to processors.
However, I have no experience at all with the MIC Xeon Phi. Has someone around here some experience with LAMMPS and the Xeon Phi ? How does the acceleration with the Xeon Phi compare to GPU ?
Another important question, is it possible to run LAMMPS using both Xeon Phi and GPU ?
Many thanks in advance !
With best regards,
Christophe
Dear all,
At my lab we are thinking of buying a new cluster (not too large) which
would combine GPUs with either processors or Xeon Phi. However, some
questions need to be answered before we decide if we buy procs or the MIC.
for the current generation xeon phi hardware, i would suggest you stay
with processors, unless you are a developer or have some mandate to use the
xeon phis. it may be worth to obtain a Phi for development purposes, but i
would wait until the next generation hardware becomes available, since it
will be operated differently.
So far I have been using LAMMPS with GPUs (K80) and I must say it works
well. I observed in general (at least for my problems) speed up factors of
about ~ x5 in comparison to processors.
hmmm... that is either not overly good, or you have a problem, where you
don't have as much potential for acceleration, or you are not using the
GPUs in the most effective way. looking at it from the cost perspective, at
5x you are only moderately better than the price ratio between the GPU
hardware and CPUs and also the power consumption benefit (i.e. running
cost) is not that much in favor of using the GPUs.
However, I have no experience at all with the MIC Xeon Phi. Has someone
around here some experience with LAMMPS and the Xeon Phi ? How does the
acceleration with the Xeon Phi compare to GPU ?
it will depend very much on the specific system and how much of your
script is accelerated as well as how far you can tolerate single precision
kernels. there is no way telling from the outside, what you will obtain.
you need to benchmark your specific use cases.
Another important question, is it possible to run LAMMPS using both Xeon
Phi and GPU ?
no.
also, i would consider it a bad idea to set up a machine, where you have to
use both at the same time. you have to deal with the complications of both
hardware architectures. it is hard enough to deal with one of them.
axel.
Ok, good to know.
What do you mean by “not using the GPUs in the most effective way” ? How can this be done with LAMMPS ?
Ok. We have a cluster with some Xeon Phi. I will do some tests.
Totally agree with you !
Thanks for all the info Axel.
Christophe