3 Oct
2024
3 Oct
'24
7:57 p.m.
On 10/3/24 14:25, Chuck Guzis via cctalk wrote:
On 10/3/24 10:51, Norman Jaffe via cctalk wrote:
Yes, this was a huge error on IBM's part. There are some
standard numerical algorithms that work by examining small
differences between numbers and then iterating. The varying
level of precision (up to 3 lost bits of significance) made
all these iterations very problematic on the 360/370. IBM
finally admitted their mistake and put IEEE floating point
on their later mainframes as a software-selectable option.
I helped one physicist who was trying to run a big VAX
astrophysics simulation on the U of MO 370 that ran into
this. It ran fin on our VAX, of course.
Jon
The IBM 360 single precision floating point has a
range of 10**-79 to 10**75; double precision and extended precision has the same number of
bits for the exponent.
...and most significanrly, normalized to only the hex digit
(4 bits),
not to the bit. Really awful if you thought you could migrate code
from a 7094...