Python on 64 bit/IA-64 ...

Tim Peters tim_one at email.msn.com
Wed May 10 06:53:22 CEST 2000


[Bill Tutt, as quoted by Trent Mick in a msg I didn't actually see
 on the newsgroup]
>> On x86 alone the only consideration really is size.
>> Speed isn't an issue, the FPU core always does all of its
> computations at long double percision. (80 bit)

Alas, that's no longer true:  on the Pentium III, the speed of fp division
is greatly affected by the setting of the FPU's "precision control" bits,
and you pay big for leaving it at the default ("long double") value.  Later
gimmicks in this architectural line are very likely to extend the speed
differences to more areas.  Double alas, since Intel stopped publishing
comprehensive timing info with the P3 release, it's darned hard work to
figure out the truth anymore.  Triple alas, while 80-bit precision is a
wonderful HW feature, the utter unpredictability of when compilers dump
temps to memory (rounding back to 64 bits as a side-effect) makes it worse
than useless for serious numeric work.  As a result of all that, it's easy
to predict that 80-bit arithmetic will get slower and slower relative to
64-bit over time:  there's alomst no customer demand for 80-bit floats, and
fast support is expensive in silicon.

hw-designers-stop-designing-what-software-refuses-to-take-seriously-ly
    y'rs  - tim







More information about the Python-list mailing list