[gmx-developers] Posit-enablement of GROMACS

[Erik Lindahl]

Hi,

On Sat, Aug 4, 2018 at 8:12 PM Theodore Omtzigt <theo at stillwater-sc.com>
wrote:

>
> It doesn't need to be low hanging fruit. We are building a custom
> supercomputer for this,
> so we are interested how to speed up this workload.
>

What Mark is saying is that MD codes are almost entirely floating-point
limited, and the last few years we have been pretty good at improving
algorithms to the point where we reach a decent fraction of the max
theoretical FLOP-rate, including single-precision SIMD instructions. Using
modern Skylake-EP x86 cores (dual avx512 FMA SIMD units) as an example,
that means the theoretical FLOP-rate of each core is 2*16*2=64 flops per
cycle, and our inner kernels can get quite close to that rate.

For any new floating-point format to be competitive, it would need to
provide similar performance per core. I haven't studied the details of
posits, but unless you can execute them e.g. with single integer
instructions, you will need a new custom chip architecture rather than just
building a computer :-)


> In broad strokes, posits improve on IEEE float in reduced complexity,
> higher accuracy,
> and adherence to associative and distributive laws of arithmetic. The
> latter is where
> the reproducibility comes from.
>

... and I predict none of that is going to be relevant at all to users if
it comes at the cost of 10x lower performance :-)

>
> If we need some education on how best to apply this to MM-style MDs, we
> are eager
> to learn how best to apply this from GROMACS SMEs.
>

While I think innovation is great in general, the caveat for us is that I
think we have experience with 5-10 esoteric/new/special architectures that
never made it into the market for various reasons ;-) So, the one thing
that would be interesting is benchmarks showing posits provide higher
absolute performance for a highly optimized floating-point limited
algorithm that already used single precision on a broadly available
architecture.

Cheers,

Erik



-- 
Erik Lindahl <erik.lindahl at dbb.su.se>
Professor of Biophysics, Dept. Biochemistry & Biophysics, Stockholm
University
Science for Life Laboratory, Box 1031, 17121 Solna, Sweden
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20180804/bc5f57c4/attachment.html>