[geeks] Yeah, the G4 is faster...

[joshua d boyd]

On Fri, Aug 17, 2001 at 08:03:36AM -0400, Ken Hansen wrote:
> In my understanding, an MFLOP is a Million FLoating point OPerations,
> correct? If the CPU cycles through, oh, let's say 500 Million clock cycles a
> second, and *if* the smallest FLoating point OPeration can be executed in
> *one* clock cycle, shouldn't it max out at 500 MFLOPS?

In the absense of SIMD and superpipelining, yes.  However, the G4 has
both.  

> By my back of the envelope calculation, you've got aprox 3 FLoating point
> OPerations going on at the same time...

I believe that it has two floating point units (just pulling stuff
from a highly volible memory bank, specifically my brain), and then an
altivec unit (which unlike mmx is able to be used at the same time as the
floating point unit).  The altivec uses 128bit registers, which means that
it can operate on 2 64bit floats, or 4 32bit floats.  So, if you sum this,
it means that the G4 can be working on from 4 to 6 numbers at once
(assuming that 16bit floats are inadequate).  That is how you get
gflops=3xMhx
 
> I don't get it.
> 
> BUT, I assume my understanding of what an MFLOP is is too simplistic, so I
> would appreciate if someone could provide me an explaination to have these
> two conflicting pieces of information make sense...

Your understanding of the MFLop was fine, it was the understanding of the
G4 that was flawed.

-- 
Joshua D. Boyd
http://www.cs.millersville.edu/~jdboyd/

I'm just here because I was in the comic book. -- Alexandra Cabot