Re: Flexible Floating-Point Standard Proposal

From: Steve (Steve_at_NOSPAM.smart-life.net)
Date: 03/16/05 

Date: Wed, 16 Mar 2005 15:18:51 -0700

John,

> This is an interesting idea. Why is the committee dealing with IEEE
> 754 ignoring it?

NIH = Not Invented Here. About a decade ago I was approached to
coordinatge the IEEE FP effort, My response "GREAT, now I fan fix all of
the problems it has!" "What Problems?" The next half hour was taken up
with my explanation of its many problems. That was the end of THAT
discussion. I probably should have just kept quiet until I had the position.

> There is actually a simple reason. Although an interesting idea, it
> would require effort to implement. Thus, the fact that it exists as an
> alternative is not enough to lead to an effort to standardize it.
>
> The effort to create IEEE 754 resulted from a groundswell of demand
> among people using computers to perform numerical calculations, based
> on the fact that the floating-point offered on at least some
> architectures (i.e., the hexadecimal floating-point of the IBM 360
> series) was inadequate, and that gradual underflow was sorely needed.
>
> Some items, such as affine infinity, were dropped from IEEE 754 - they
> were implemented on the 8087, but not on its successors, therefore.
>
> If enough people who work with numbers think that something like your
> proposal is needed, then, once they take it up, this idea *will* go
> somewhere; but there has to be a real demand before it is felt that it
> is worth the effort.

Here is the problem. Global Warming, outsourcing, the balance of trade,
etc., are all things that are now being allowed because present
simulations completely and hopelessly fail to predict their effects.
Further, there is good reason that these fixes to FP would make much
more reliable simulations possible. However, the people doing this don't
represent 1% of the PC users. Seeing some of the logical traps that the
VERY competent people on this forum fall into (e.g. using algorithms
that don't preserve significance/dimensionality) and complain about
non-existent problems, what chance is there for the community of users
to ever understand these issues.

Indeed, much of the present '754 is via "proof by authority" of the
person now commonly referred to as the "Father of IEEE FP", when in fact
even HE lacked the skills to do a good job of this, to the point of
leaving no smooth path for extension.

As I see it, the fundamental process is broken with little prospect for
broad-spectrum repair. Even PhD /Math people lack the skills needed to
participate usefully in this discussion. The ONLY hope I see is for some
competent person in Sun or Intel to "see the light" and run with this.

Steve Richfie1d