[Beowulf] Moores Law is dying
evm at stccmop.org
Thu Apr 9 11:04:53 PDT 2009
Moore's Law was the observation that the number of transistors per die
that was economical to produce doubles in about 18 months. As process
technology advanced our yield increased and transistors shrank. Higher
yield only gives factors of 2 to 10 in the number of good die (this is
the economical part) and is a battle fought with each new process
innovation. Transistor size has a hard limit in 2-D which is reached
when we approach atom scale devices. I imagine that Front end gates with
oxides 3 atoms deep might just be very hard to make economically. This
then brings us to either (or) going 3-D which can give us a few Moore's
Law cycles (2 layers, 4 layers ) or making the die larger. Making the
die larger has been done in a number of instances. Tektronix's 2" square
backsided imager comes to mind. I'm equating the thinning process for a
backside device as the same in complexity as going from a ccd imager to
a multi core cpu.
I can well imagine a 1024 core device, and 4 of those in a 1U box. Intel
has an 80 core device in the lab. Sure there are heat, power, memory
problems to overcome but that is the name of the game.
We have seen the power of CPU grow such that even windows is not able to
bring the faster processors quite to their knees. Thus the drive to the
Atom and its like. I can imagine a notebook with a 32 core cpu that has
slower cores than the atom. Might just be a screamer.
The demise of Moore's Law in 10 to 15 years may happen. That is if you
use the law as stated (transistor count). If you consider the power
(MIPS,Flops etc) if a system then clustering waferscale devices gives us
Moore's Law growth potential for the next 20 years. Plus there are new
technologies to be discovered.
For me the demise of Moore's law is a function of economics. How will
Intel get its money if the cost per core is $1 and 8 cores more than
covers your desktop (1,000,000 notebooks with $32 cpu does not pay for a
new fab). Making the Waferscale 1024 core cpu takes a lot of research
money. Will the sales support it?
Ethan VanMatre - Research Systems Engineer - Computer Systems Manager
Center for Coastal Margin Observation and Prediction (CMOP)
Oregon Health & Science University
(503) 748-1157 evm at stccmop.org
More information about the Beowulf