----- Original Message -----
From: "G.F. Thomas" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Monday, December 15, 2003 7:27 PM
Subject: Re: Fortran 95 on Opteron
> ----- Original Message -----
> From: "Aleksandar Donev" <[log in to unmask]>
> To: <[log in to unmask]>
> Sent: Monday, December 15, 2003 4:29 PM
> Subject: Re: Fortran 95 on Opteron
>
> > What exactly does 64-bit mode get you for number-crunching
programs?
>
> Less than you expect really. Sure you have more registers and a
vastly
> increased addressable space. You'll be able to run bigger problems
> faster but not necessarily with any greater accuracy. If it's a
> commodity processor like Intel or AMD the fp calculations may be
done
> in extended precision but you'll have little or no control over that
> unless your compiler supports the extended precision type which few
> do. Some compilers support the long double type which is preferable
> whether in 32- or 64-bit (addressable space) mode.
>
Note that a valid program run on the same IA-like processor in 32- and
64-bit modes will generally not give output that is identical to1 ulp
although they should be close. In IEEE-speak, the two modes are
different destinations and are expected to differ in minor and
insignificant ways. Furthermore, in 32- and 64-bit mode, single and
double reals have the same widths (4 and 8 bytes, respectively, for
IA-type processors) and the same precision & exponent widths
(including the sign bit, 23 & 9 =32 and 53 & 11 =64 bits,
respectively). For extended double fp register arithmetic in 32-bit
mode, the precision & exponent widths are 64 & 16 =80 bits; I've
forgotten what these widths are in 64-bit mode but I recall that they
are different for Intel and AMD. Add to the foregoing that the
rounding behaviors in 32- and 64-bit modes can differ.
--
Ciao,
Gerry T.
______
"In a world in which the price of calculation continues to decrease
rapidly, but the price of theorem proving continues to hold steady or
increase, elementary economics indicates that we ought to spend a
larger and larger fraction of our time on calculation." -- J.W. Tukey.
|