Andy Eng said:
I'm familiar with custom hybrids with parallel processing dies and must
confess to not having kept up with the industry over the past decade --
That both Intel and AMD are now mass producing duals in a single
package, are the factories bumping up against Moore's law or are these
devices simply taking advantage of available packaging advances?
As I understand it, there are some interesting gotchas which are
currently affecting the applicability of Moore's Law to processor
design.
As you increase the number of semiconductor junctions per square, you
end up having to use even-finer-pitch deep-submicron fabrication
processes, thinner gate insulators, and lower voltages. Parasitic
effects (inductance and resistance of the metal layers, capacitance
between nearby traces and gates) begin to have a significant effect on
the circuit performance.
Leakage current across the gates and through the transistors starts to
climb, so that even a circuit which isn't being clocked is drawing
power and dissipating heat. This can necessitate the use of
variable-voltage circuitry (parts of the chips which can run at lower
speed are fed lower voltages), or smart-powerdown logic, etc.
Total power supply currents become quite large (many tens of amperes
at 1-2 volts) and the high currents lead to higher I*2*R losses.
Power densities on the chip increase, leading to a need to remove a
great deal of heat, very efficiently, from a very small area. I've
seen pictures of what happens (very quickly) if an Athlon CPU loses
its heatsink... smoke and flames within a few seconds. There was one
famous comment made not all that long ago, that if power densities
continued to climb at their current rate, it wouldn't be many years
before a high-end CPU has a power dissipation rate equivalent to a
similar-sized patch of the sun's surface!
What it boils down to, is that in some situations, you can get a
certain amount of total CPU performance less expensively by using two
or more CPUs or cores, clocked at a lower speed, in place of a single
CPU/core clocked at a higher speed. As long as the operating system
is capable of taking good advantage of multiple CPUs, money can be
saved.
I saw one embedded-system board described this week in a trade rag,
which uses a pair of 1-gig VIA processors rather than a single
higher-speed CPU. It's capable of running with passive/convective
cooling only - no CPU fans.
So, I think the answer to your question as to whether it's improved
packaging / expertise, or problems in taking advantage of Moore's
Observation, is that it's both.
