I just came back from the reception for people who won such-and-such a fellowship, where I got a great chance to talk one-on-one with the UR alumnus that runs Intel's high-end server division. He led the group/was responsible for figuring out how to change the voltage/frequency of a chip on the fly. 40 minutes or so, let me throw down what I remember - I wasn't exactly prepared to take notes.
-Intel creates a new chip fabrication process every two years, at a cost of $3billion or so. It used to be every 3-4 years, but it's been consciously accelerated. The business is cutthroat.
-Most of their research is fabrication. They only have one small lab in Oregon doing circuits. The quality of the fabrication process is responsible for about three-quarters of the cost/quality of a microprocessor.
-Intel and AMD have completely different production processes. Intel's per-unit production cost is significantly less than AMD's. He claims their method is more reliable, too, but the technical details escape me. AMD is apparently still learning, they pay IBM lots for technical expertise.
-Within five years, they estimate that Google will purchase one-third of all microprocessors. Google already exerts significant monopsony power in the market, since when they construct a new standard build, the chips have to come either exclusively from AMD or Intel. AMD's last few profitable quarters a couple years back can be traced to their holding of the Google contract.
-Large scale purchasers like Google, eBay, Amazon, don't buy high end product - just tons of the basic units, what anyone would have in their homes. Margins on servers are much, much higher than on microprocesors.
-4 years of nonstop running of a modern Intel processor will result in a .1% chance of failure.
-Moore's Law still has several years to run yet.
Think that's the basics of it. Was a great chat.