IntelUser2000 (Intel_user2000@yahoo.ca) on 5/16/07 wrote:
>
>BTW, why would they take the legacy FPU to 128-bit anyway??
>I think they are still 64-bits, exactly similar to Yonah.

Legacy FP is 80 bits, not 64. And I believe that it is also
usually expanded by a few bits internally, in order to be
able to do 80-bit complex math (trig etc) more easily and
get all the nasty results right within one ULP even though
you have intermediate results.

So I think the legacy FP path tends to be on the order of
82 bits or so, not 64.

But yeah, I don't see any reason to do a 128-bit path,
other than in specific cases (ie the multiplier would
need to have 128-bit internal paths at some point, since
the mantissa in a 80-bit FP value is 64 bits).

Of course, you could probably make the case that nobody
actually cares about legacy 80-bit FP any more,
and not even implement it. I think that's what all the
x86 software emulators do (except on Itanium, which I
think has full hw support for the 80-bit extended math).

But I don't see Intel or AMD throwing out their existing
support for 80-bit math.

Or did you mean something else?

Linus