Page 164 - A Practical Guide from Design Planning to Manufacturing

Microarchitecture 137

may not be practical. An important area of research is finding ways to
beat Pollack’s rule in order to continue to improve performance but with
smaller increases in die area.

Measuring Performance
Processor performance can determine easily whether a computer prod-
uct is a commercial success or a failure, and yet it is frustratingly hard
to measure. Imagine the performance of several cars must be rated, giving
each one a single number as a measure of its performance. Top speed could
be used as a measure of performance, but consumers would likely com-
plain that they don’t often have the opportunity to drive at over a 100
miles per hour (mph). The lowest time accelerating from 0 to 60 mph
might be a more meaningful measure of speed drivers would care about.
Alternatively all the cars could be driven through an obstacle course and
their times compared to see which was the fastest under more realistic
conditions. Which car rated as the fastest could easily turn out differently
depending on which measure was chosen, and the manufacturers of the
other cars would inevitably complain that this rating system was unfair.
Measuring processor performance has the same types of problems.
One of the earliest metrics of computer performance was determined
by measuring millions of instructions per second (MIPS). The MIPS
rating was simple and easy to understand. Unfortunately not all instruc-
tions require the same amount of computation. A million adds in one
second are very different from a million branches. When DEC launched
the VAX-11/780 computer in 1977, it was sold as a “1 MIPS” computer
because it had similar performance to an IBM computer that was being
marketed as a “1 MIPS” machine. The VAX computer was so popular
comparisons to it were used to determine the MIPS of other computers.
It was only in 1981 that tests showed the actual rate of instruction com-
pletion on the VAX machine to be about 0.5 MIPS. 3
Had DEC been lying all along? In terms of the absolute rate of execu-
tion, the VAX computer did not run at 1 MIPS, but its more complicated
instruction set let it accomplish the same work in fewer instructions. It
could run programs in the same length of time while executing instruc-
tions at a slower rate, so it was a 1 MIPS machine in terms of relative
performance. Eventually MIPS came to mean how much faster a com-
puter was than the VAX-11/780. Because of the problems in comparing
computers with different instruction sets, some in industry came to joke
that MIPS really stood for “Meaningless Indicator of Processor Speed.”

3
Hennessy and Patterson, Computer Architecture, 72.