[gmx-developers] Re: Posit-enablement of GROMACS

[Theodore Omtzigt]

On Wed, Aug 1, 2018, 16:12 Mark Abraham <mark.j.abraham at gmail.com> wrote:

> It does sound like a generic hope, rather than one based on extensive
> understanding of how high performance MD codes work. :-) I don't mean to
be
> negative, but MD codes are probably not the low hanging fruit for
> demonstrating the benefits of this approach.

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.

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.

The basic mechanism to produce reproducibility is by explicit management of
rounding
events. In a posit system, we can accumulate the unrounded output of the
adders
and multipliers and if you say, 1/sqrt(x) is important we can accumulate
that unrounded too,
although we all know that is not quite a true statement when div is
involved.

Our initial thought was that this mechanism to accumulate large number of
arbitrary small
and large values without error could be used advantageously for the force
field calculation.

Secondly, for forward/inverse transform calculations, this mechanism yields
a pure identity
matrix, and our FFT/iFFT take advantage of this and suddenly 16-bit posits
can compete
with 64-bit doubles.

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.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20180804/8c63d8aa/attachment.html>