As some of you may already know: Intel now distributes its compilers for
free:
https://software.intel.com/content/www/us/en/develop/tools/oneapi/all-toolkits.html#gs.4siqji
There is an EULA, but I don't see any gotchas in it, not being a lawyer.
You will need both the Base toolkit and the HPC toolkit. Some preliminary
notes based on my (limited experience):
1. Amber things work for me, both on Linux and OSX (Intel CPU). I can't
compile cuda with the Intel compilers, since my nvidia toolkit doesn't
think it is compatible. (I've only tried cuda 10, not yet cuda 11). CPU
MPI using Intel's MPI stack works, and MKL integrates correctly.
2. There are cases with significant speedups using the Intel compilers,
compared to GCC. I see them in RISM and xray stuff, but that is what I
have been working on. Intel compilers appear to be quite good at
implementing OpenMP directives.
3. The Intel python installation also works fine. It seems to have
all the packages Amber needs. Ditching miniconda has not given me any problems;
just point run_cmake to your Intel python executable. I have no clue about
whether there is any performance enchancement or not, but avoiding miniconda
downloads has some upsides.
Don't switch just for the sake of switching. But I found the downloads and
installation to be pretty simple, and the former hassle of purchasing a
license (or lying about what you are using the software for) is gone. And
since users will likely being doing this, at least some developers ought to
be testing with Intel compilers. If nothing else, you'll get some new
compiler warnings....
....dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Jun 25 2021 - 12:00:03 PDT
