Improving Hardware and Software Reply
Last updated at 12:35 am UTC on 17 January 2006
Bob Jarvis wrote:
>That's an interesting question. I'd like to ask two related questions:
> Why and how does hardware get faster?
Hardware (specifically, the microprocessor) has achieved a factor of
roughly 500x performance improvement over the past 20 years. By far the
largest factor in this is raw clock frequency, going from ~16MHz to
greater than 1GHz (60x). Most of this is directly due to process
improvements ("process" here means the feature size and series of steps
used in fabricating devices on the silicon wafer).
As more and more transistors were able to be packed into the same area,
due to the same process improvements that gave us higher frequencies,
microarchitectural improvements were able to achieve another factor of
10x. These improvements, such as pipelining, superscalar and speculative
execution, exploit various levels of parallelism, and trade off increased
die area for decreased time. Larger caches also exploit spatial and
temporal locality found in almost all code.
The demand for high-performance processors has been large. Since the
cost to process a silicon wafer is relatively fixed, shrinking the die
size via process improvements allows many more parts to be produced for
the same fixed cost. And as a side-effect, the parts run faster (since
those same process improvements give us higher frequencies). Processor
designers then have a new process available in which to pack more
performance-enhancing features in the next-generation processor, and we
jump on the treadmill again.
Another factor here is that improving performance is a fairly easy goal
to quantify and verify. It can be attacked by standard scientific and
> Can we do similar things to improve software?
Depends upon what you mean by "improve". There's no economic incentive
to reduce the size of software, since replication is essentially free.
And smaller software doesn't automatically run faster, anyway.
Software users complain more about general software "bugginess" rather
than performance. But program correctness is a much harder goal to
quantify and verify.