Bill Dally: Life After Moore's Law

In this case, the author confounds the real Moore's Law, which states that in a certain time period---1 year in 1965, 2 years later---the number of transistors of the lowest cost you can put on a die will roughly double with the resulting "speed".

Amazingly, Moore's Law is still holding up, but not the implicit promise of increased speed:

There was a corollary of generally increased clock speeds as everything got smaller, but as noted in the article we hit a brick wall there due to power dissipation (e.g. see the Netburst (P4) microarchitecture disaster).

But from my informal understanding of CPU speed, the greatest gains have come from better microarchitectures. There are fairly clean examples of this in e.g. the evolution of the 680x0 and 386 and beyond lines of microprocessors, and we can see this today in examples of x86-64 and ARM processors that are superscaler (or not?) and in or out-of-order (Atom vs. most everything else, Cortex A8 vs A9, etc.)

Our problem here is of course that the microarchitects have run out of the big wins, e.g. "Per Core, clock-for-clock, Nehalem provides a 15–20% increase in performance compared to Penryn" (http://en.wikipedia.org/wiki/Nehalem_%28microarchitecture%29...). Nice, but not a big win, not like the 386 -> 486 -> P5 -> P6 steps, where among many other things cache was added, it went superscaler and it went out-of-order superscaler.

Maybe the author gets this but it didn't emerge in the final draft, but I have no way of knowing that....