[Info-vax] OT: Computing Experience, What brought you to VMS?

[glen herrmannsfeldt]

already5chosen at yahoo.com wrote:

(snip, I wrote)
>> The 11/730 had H-float as standard, it was an optional extra
>> for most other models, including the 11/750.
 
> H_floating is quad-precision. Nice to have, but not really necessary 
> for ordinary uses.

Well, mostly because numerical analysts have found enough ways to
work around double precision.

Well, the primary use for double precision is to get enough bits
through the intermediate computations to get single precision
results. 
 
> In the statement above I had in mind G_floating - the most 
> universally useful among VAX floating-point formats.
> I was under impression that G_floating was available on all 
> VAX models. Is it incorrect?

As I understand it, on some you had the choice of either D or G,
and optional extra charge for H. Or, I believe, optional extra
charge for both D and G.  (I am not sure when you run into the
limit on microcode size.)
 
>> > On the other hand, CDC floating point arithmetic had reputation 
>> > for very bad numeric properties. Was it fixed in Cyber 200 series?
 
>> Makes sense if you consider what Cray did on the later machines.
>> There are many algorithms where close is good enough. If you are
>> a little careful, the deviations will average out, and many of those
>> need a fast processor.
 
> I disagree.
> Of course, fast, precise and consistent is the best. But if I 
> would be pushed to give up on one of the three I'd rather give 
> up on the 2nd, then on the first, but not on the 3rd.

Then I presume you don't work on those kinds of algorithms.

Among such are the iterative solutions to partial differential
equations, especially in 3D. Some of the uses for such are
weather forcasting and classified simulations done by DOE labs.

Or you might just look at the Newton-Raphson square root algorithm.
It is usual to do the initial cycles in single precision, with the
last ones in double precision. More specifically, with a good initial
approximation and dividing the exponent by two, two cycles in single
and two in double precision will do. Doing the first cycles in 
double precision doesn't change the result. 

Unless you do a huge number of square roots you won't notice,
but for a weather simulation it might make the difference 
between taking half a day or two days to compute a forecast
for tomorrow's weather. 

-- glen