If there's one thing that gets my goat, it's bad math and
statistics.
In Gauging
the Gigaflops Gap, Don Granberry of ZDNet tries to explain how the
G4 with AltiVec is superior to the Pentium III. It's actually a fairly
accessible explanation, but the numbers are flawed.
Granberry uses the following table to compare a 700 MHz PIII with a
500 MHz G4:
It's a nice table. It's clearly labeled. The facts are clearly
established. But there's one piece of misleading data: CPU speed
doesn't mean a thing when you're looking at how many clock cycles are
used.
MHz for MHz, the G4 is 3.68 times more efficient at these particular
functions (or, as Granberry puts it, the G4 does the same work in 27%
as many processor cycles).
This wouldn't be a problem, except that later in the article he
links clock speed to this. Which wouldn't be a problem, except that he
doesn't apply the same clock speed to both processors.
Multiplying 3.68 by 700 MHz, Granberry states that a 2.6 GHz Pentium
III would match the performance of a G4/500 on these tests - but it
just isn't so.
The problem is the misleading data (remember how story problems
would sometimes introduce numbers that were not necessary to solve the
problem?) brought him to the wrong number. No matter how precise the
math, if you're multiplying the wrong number, you get the wrong
result.
Instead of 2,576 MHz, he should have multiplied 3.68 by 500 MHz,
resulting in a Pentium III running at 1,840 MHz to match the G4/500's
performance on these particular functions.
Of course, Apple is now shipping the Power Mac G4/500 in a dual-processor
configuration, so it would take 3,680 Pentium III MHz to match it. For
the sake of convenience, let's say a quad-processor PIII/900 system
would match the dual-processor G4/500 on these vector functions.
Whether this means much in the real world is another question
entirely, but at least we have the math right.
