Sun's Greg Papadopoulos, the man behind the Galaxy servers and much of Sun's x86 push, put a fascinating essay on the continued operation of Moore's law as microsystems replace microprocessors in his Sun blog last week.
Here are some key paragraphs:
Just as the '80's discrete processors were killed by microprocessors, today's discrete systems --- motherboards full of supporting chip sets and PCI slots with sockets for microprocessors --- will be killed by microsystems: my word for the just-starting revolution of server-on-a-chip. What's that? Pretty much what it sounds like. Almost the entire server (sans DRAM) is reduced to a single chip (or a small number of co-designed ones, just as the first micros often had an outboard MMU and/or FPU). These microsystems directly connect to DRAM and to very high speed serialized I/O that are converted to either packet or coherent-memory style network connections.
Open up the lid of a microsystem and you'll find a full SMP : multiple processor cores, crossbar switches, multi-level caches, DRAM and I/O controllers. Our Niagara chip, for example, has eight cores (each four-way threaded), a caching crossbar switch, four memory controllers, and a high speed I/O channel. And its performance is very competitive with our original E10K, the 64 processor behemoth that stormed the world as the first high-volume, enterprise class, massive multiprocessor.
The secret was to turn the clock back --- figuratively and literally --- to earlier, more sane processor pipeline designs. Ones that were more conservative of transistors, area, power, and complexity. (A key innovation, however, was to finally fold multithreading into the individual pipes). With these smaller, leaner and far more power-efficient processor cores, we could then use the transistor count advance of Moore's Law to paste down many of them on the same die, and to integrate the rest of the SMP guts at the same time.
Where does end up? Well, we are now dying to get to 65nm (Niagara is 90nm) so we can get even more transistors on a chip in order to integrate more and bigger systems. Just as the microprocessor, harvested the pipeline inventions of 60's and 70's, microsystems are going to integrate the system innovations of the 80's and 90's.
By 2010 microprocessors will seem like really old ideas. Motherboards will end up in museum collections. And the whole ecology that we have around so-called industry standard systems will collapse as it becomes increasingly obvious that the only place that computer design actually happens is by those who are designing chips. Everything downstream is just sheet metal. The apparent diversity of computer manufactures is a shattered illusion. In 2010, if you can't craft silicon, you can't add value to computer systems. You'd be about as innovative as a company in the 90's who couldn't design a printed circuit board.
Sun's Microsystems offer Solaris/SMP on a chip, Microsoft has discovered threading and security on the PPC model, and IBM has its grid on a chip ready for Linux, but where's Apple? If Papadopoulos is right, and he is, when he says "In 2010, if you can't craft silicon, you can't add value to computer systems" he's also telling Steve Jobs to either get out of the computer business or get his head straight on CPU issues.