🚀 go-pugleaf

plaw@usc.edu (Patrick Law) writes:

>Correct me if I am wrong. For the last 20 years, we have 4-bit 
>microprocessor, CISC, cache, RISC, superscalar, superpipeline,
>etc. Have we reached a point of diminishing return in uniprocessor
>design? Could we still get a speedup of "X" every "N" years like
>before? The above evolutions are just experiences we learnt in
>building big machines like the IBM 390. I think it is a time for
>MIMD and a change of serial programming style.

That has been the opinion of many people for many years. Why then, has
it not happened? My guess is that there are two reasons:

1) Micro-parallelism (superscalarity etc.) in uniprocessors has kept
them going faster at a steady rate for many years.

2) Implementing "traditional" programming styles on MIMD architectures
has met with little success. Languages that support parallel
programming style has not gained widespread acceptance for
"mainstream" programming.

The real questions are then

ad. 1) How long can we expect this to continue.

ad. 2) Can we expect any change is this area.

My guess is that uniprocessors will continue to increase their speed
at approximately the present rate by use of new fabrication
technologies and extended micro-parallelism. The relative cost will
however rise faster than the speed. We have already passed the point
where multiprocessors based on cheap processors offer more raw
performance per $ than fast uniprocessors. The programming problems of
the multiprocessors has kept these from gaining popularity.

This brings us directly to the other question. I fear that we will
never get acceptable compilers for low-level languages (like C) for
MIMD machines. So, as long as people insist on using these languages,
they (and as a consequence, we) are stuck with uniprocessors. People
need a very good incentive for changing to a completely different
programming language and style, and the only good incentive for the
people who matter (the industry) is money. So until someone produces a
parallel machine that runs the programs people want at twice the speed
and half the cost of competing uniprocessor designs, there is not
likely to be any change.

Specialist areas will in all likelihood use more parallel machines,
but these tend to be very expensive systems, either because they
combine a huge number of processors (like CM2) or because they are
based on (relatively few) expensive high-end uniprocessors (like CM5).
These systems are not likely to compete against workstations and
high-end PCs.

	Torben Mogensen (torbenm@diku.dk)