Re: Hardware?
On May 16 2011, Corliss, George wrote:
On Kahan's website you'll find one presentation that makes this point: HW
vendors focus, for good or bad, on performance benchmarks.
http://www.cs.berkeley.edu/~wkahan/ieee754status/baleful.pdf
Well, yes - he wasn't the first or last person to make those points.
They were an objection to the original Linpack test suites, even before
the vendors picked them up.
IMHO, what P1788 should be doing is this:
* Develop a spec about a compliant HW/SW system - leave mandatory HW
details out.
Or even produce a draft! I may have missed that but have asked a couple
of times, and it's impossible to know the background without such a thing.
But you are right that such a specification should not mandate HOW it is
implemented - I have used Fortran compilers with software floating-point,
and that was conforming.
* Develop interval arithmetic applications that a broad spectrum of
users want to run, and run fast where is that killer application
(interval spreadsheet, cell phone or tablet app?)
You will first have to demonstrate that there ARE such applications!
Getting support into Matlab/Octave/whatever would be a very good start;
some of us could use that for teaching even if a relatively small amount
of such functionality were interval-enabled.
But a bigger issue would be to demonstrate that it is viable for a wide
enough range of uses. In particular, showing that the problem of
convergent iteration appearing divergent is soluble (if it is!) would
be critical. Even the numerically competent subset of the outside world
remains to be convinced about interval arithmetic's usefulness; most of
the remainder have never heard of it.
* Get those interval arithmetic applications incorporated into benchmarks
that are used to evaluate processor performance.
Until and unless more people are convinced of its usefulness, I doubt the
feasibility of that.
Has P1788 thought of creating a C/C++/JVM (pick your language) library
that complies with the P1788?
That would be a good start to getting it into applications, especially
if the language were C++ or (preferably) C.
Regards,
Nick Maclaren.