Intel Phi speed impressions?

LAMMPS LAMMPS Mailing List Mirror
1

--- clusters ----
I know XSEDE's Stampede uses Intel Phi coprocessors.

Alas, no benchmarks yet.
http://lammps.sandia.gov/bench.html

Has anyone tried running LAMMPS with USER-INTEL or KOKKOS?

  --- desktop machines ---

I have been asked for suggestions to by a new workstation and I am
generally curious whether or not it makes any sense to buy and set up
this kind of specialized hardware on a desktop?

To be slightly less vague: my own usage scenario would be a
(coarse-grained) system with 10^5-10^6 Lennard Jones particles
(without long-range electrostatics, but I imagine other users on the
mailing list might be curious about long-range electrostatics
performance as well). In our case, we probably want to keep the total
cost well below $8000.

Any suggestions would be welcome, even if they have nothing to do with
Intel Phi.

Thanks!

Andrew

2

  --- clusters ----
I know XSEDE's Stampede uses Intel Phi coprocessors.

Alas, no benchmarks yet.
http://lammps.sandia.gov/bench.html

Has anyone tried running LAMMPS with USER-INTEL or KOKKOS?

oops. Vague. My apologies. I meant with Intel Phi coprocessors.

3

Hi Andrew,

4

Thank you for your thorough answer.

Andrew
(Your workstation (i7+GTX980) is approximately what we can afford to
buy, so this information is perfect for me.)

5

Hi Andrew,

Most of the focus for the Intel package thus far has been for soft matter with long-range electrostatics, although this is changing... Here are a few benchmarks that include Stampede (which is nice for comparisons of GPU vs Xeon Phi in parallel):

https://sites.google.com/site/wmbrown85/brown_hpcuf_54.pdf

slides 17, 19, 23, and 24. The missing numbers on 24 (hw unavailable at time) are 2.84X and 9.29X respectively.

I typically only work with server/cluster/hpc parts - (that have ECC, higher core counts, etc).

If you want to send an input script to me directly, I can take a look and get back to you. For simulating with few atoms on many-core there might also be some options to improve overall throughput that I believe could work well with Phi.

Also if you want, I can add you to my announcement list for changes in the Intel package for LAMMPS (although it is probably just as easy to check bug fix page).

Best, - Mike

6

Dear Michael
    Thank you for the link. This is helpful. I will forward this to
several people.
Andrew