aglew@ccvaxa.UUCP (02/27/88)
LARGE DATA SUMMARY
------------------
A while back I asked a question about people who might use extra precision
floating point (that was as fast as single precision), promising to
summarize mail responses to the net. Here's my summary - note that I do
not believe in posting the raw responses.
(1) I must have been quite vague, since there were a lot of people who
said "Sure, I'll use 64 bit double precision - would never use 32 bit".
I had meant to ask the question about 128 bit precision, or, conversely,
would you buy a machine that did 64 bit precision twice as fast as
X, given that it did single precision 10% slower than X, and cost the same.
(2) Extra precision:
14 people said that 64 bits was as large as they would ever use.
8 people described applications where they would like to use >64
bits of precision.
1 of these seemed to be in a position to greatly influence
a supercomputer sale (BUY/NO BUY decisionmaker)
2 seemed to be regular supercomputer industrial or DoD
supercomputer users, with some influence.
1 said that he would use Cray double (2*64) if it were
anywhere near in performance
Several people who seemed to be students said "MORE PRECISION.
"MORE, MORE...", and wanted >128 bits.
1 was currently using a machine with 96 bit precision, and
indicated that it was the only machine he was satisfied with.
(3) "Would you buy a machine that did 2x DOUBLE if only 0.90x SINGLE?"
The overwhelming answer was no... but then, I am not interested in
the overwhelming answer, only the presence and size of the minority.
Only 4 people said that they would, 1 of whom also expressed an
interest in extra precision (and seemed to be in a position to
influence a sale)
Many people said that they wanted faster DOUBLE, but no tradeoff for
SINGLE.
(4) Other
Sorry if I have misrepresented your views, or completely ommitted
them - I lost a bit of mail.
Thanks to people like Mark Johnson of MIPSco, who sent me letters
describing design of a balanced floating point system, and to the
several others.
(5) Conclusions
I don't really know what you can conclude from the above.
The net is not exactly a scientific polling ground.
Most people don't believe in extra precision or trading off
double speed for single - but some do. One or two seemed to
be able to influence machine purchases, but the number was
not enough that I would be willing to say "Look at this great
big niche market for extra precision and speed that we are
missing".
Note that I did not try to summarize that which was sent to the net,
only what I had received by mail.