(j3.2006) Is Moore's law over?

[Van Snyder]

For a long time, software managers have depended upon Moore's law to
avoid re-thinking software to improve its performance.

During the last two years or so, processor clock speeds have actually
decreased, but overall processor performance has continued to improve,
presumably due to better microcode and fancier pipelines.

Can this continue, or is Moore's law over?

My program runs on a 362-Pentium cluster, typically taking 15 hours to
process each day's data.  Benchmarks with four-year-newer equipment
suggest we could get by with 42 nodes, each with two quad-core Pentia,
which works out to 336 cores, i.e., about the same.

Our next instrument will return 400 times as much data.

We spend about 10% of our cycles doing linear algebra, mostly inner
products but some of those are used for Cholesky factoring, which could
be done better.

We spend about 8% of our cycles doing FFT.

We're thinking about DSP/GPU coprocessors for these two subproblems, but
then there's the rest....

There isn't much of our program of the form "do n**3 things on n**2
data."  Mostly (82% or so) it's "do n things on n data."  That is, it's
mostly memory-bandwidth limited, not cache- or processor-bandwidth
limited.

Will Moore's law save us, or do we need to re-think our program?

I've argued we need to re-think our program.  We've requested money from
NASA to start investigating the problem and what needs to be done.  Some
of the guys on the review committee have said "Why should we give you
money for software development if Moore's law will save you?"  I don't
think it's wise to believe them uncritically.

Will Moore's law save us, or do we need to re-think our program?

Do you have citations from literature or web pages?

-- 
Van Snyder                    |  What fraction of Americans believe 
Van.Snyder at jpl.nasa.gov       |  Wrestling is real and NASA is fake?
Any alleged opinions are my own and have not been approved or
disapproved by JPL, CalTech, NASA, the President, or anybody else.