questions and comments about gradual underflow

[Keith Bierman ADT/QED]

>> I was wondering if you agree with my arguments above?

Not particularly. Later Alphas do better at this than early
alphas. Some of the AMD chips do better than the Intel chips. If
anything, it would seem that there exists a class of chip architect
that has decided to tackle the problem by doing better, rather than
punting on more cases.


As for the ISV's, most don't want to think about such issues at
all. So they'll take whatever the default is and complain if the
vendor is lucky, and switch to another vendor if unlucky.

Of those that care, a reasonable fraction would like to have
correct answers (that is, have the system do the best it can, so they
can worry about other things). Another reasonable fraction will just
want it to go fast (having picked their platform for some other
reason). 

Over time, I expect that systems that have egregious surprising
slowdowns will be driven from many markets. ISV's have lots of more
interesting and financially lucrative things to do than to worry about
how to tickle their vendors system, especially if it means they have
to become numerical experts to split the difference between
correctness and speed. 

Of the ISV's that *do* wish to spend time on numerics, it's possible
that they'll spend their time on interval methods, or something else
slightly orthogonal to the question at hand (however, if "fast" means
containment is lost, this customer segment will consider that choice a
non-starter).

I have no doubt that there is A market for faster but wronger
answers. The Cray market is non-zero (how much larger than zero it is
seems speculative, it isn't large enough to support multiple vendors,
and the jury is out whether it's big enough to support even 1
vendor). 

Of the current users of SPARC nonstandard modes, I'd estimate that
less than 1 in 100 have any clue as to what it means. If they
understood, I'd bet they'd be less likely to employ it (alas, they
might reconsider their platform in the process).