[PD] 64 bits of a/v pleasure?

[Mathieu Bouchard]

On Wed, 24 Dec 2003, thewade wrote:

> I thought that the processor could move more per cycle, like the
> difference between the 486 sx and dx.

The difference between the 486sx and the 486dx was that the dx had an
embedded FPU similar to the 387, while the 486sx didn't.

However, the difference between the 386sx and the 386dx was, well,
different, but marketing reused the same sx/dx labels. The dx was the
"real" 386, with a 32-bit data bus and 32-bit address bus, while the sx
had a 286-compatible box, so a 16-bit data bus.

> > > The scoop at tomshardware maked amd look really good, in 64 bit mode,
> > > and I know there are writeins for linux 2.6 on AMD 64FX...
> http://www.tomshardware.com/cpu/20030923/athlon_64-22.html#opengl_benchmarks

I doubt Quake was really in "64 bit mode". I think it ran simply in 32 bit
mode, but that this CPU handles 32 bit faster, just like all preceding
ones. Eg: a 486SX-16 was about the same as a 386DX-33; a Pentium-66 was
about as fast as a 486DX-120.

> Doesnt the ability to move 64 bits into a register and perform
> manipulations to each set of 32 bits whilst there, improve on moving
> 32 bits into registers, processing, moving out, moving another 32 bits
> in, processing, moving out?

Performing the same thing twice on two halves of a 64-bit register, is the
kind of thing you find in MMX. Performing two things on two different
32-bit registers is the kind of thing you find in all Pentiums. (In both
cases, though, there are a lot of limitations on how it can be done).

OTOH, the K8's 64-bit is about this:

 * working on 64-bit integers as you would on 32-bit integers
   (not something you'd really care about in Pd...)

 * busting the 4-gig memory limit (having 64-bit pointers).

The actual data bus of the CPU has been for many years independent of
that, with all those caches and such. The Pentium-66 already had a 64-bit
data bus although it was a 32-bit CPU. If I am reading correctly, the K8
has a 1-bit bus but with a mad clockrate: 3.2 gigabytes/second would mean
at least 3.2*8 = 25.6 GHz on a serial port (!), if I understand correctly.

________________________________________________________________
Mathieu Bouchard                       http://artengine.ca/matju