In this video from Hot Chips 25, Dr. Robert Colwell of DARPA delivers an insightful overview on the “looming issue” that is the death of Moore’s law. Colwell starts out with the humorous point that “physics doesn’t care what we believe.” Believing that Moore’s law – Gordon Moore’s 1965 observation that a chip’s transistor count doubles every 18-24 months – will last indefinitely, does not make it so.
According to Colwell, the exponential advances that have undergirded the microprocessor industry for the past five decades will start to break down over the next decade. More specifically, he expects it will come to an end between 2020-2022 at the 5-7 nanometer node. The safest thing to do, in Colwell’s opinion, is to keep pushing the envelope on miniaturization, but at the same time plan for the future.
The DARPA rep starts with three givens:
1. All exponentials must end (or they will “eat the universe”)
2. Electronics existed before Moore’s law
3. There will be electronics after it
Getting specific with the mission at hand, Colwell lays out the “canonical Moore’s law chip designer’s algorithm” as follows:
Given constraints
{Si area, schedule, team, tools, power}
How can I best design a chip to
sell huge unit volumes
at high yields
and high profits?
Colwell is doubtful that there will be a replacement that can compete with Moore’s law, which he deems the “best tech exponential improvement trend ever.” Silicon improvements have essentially been free to designers. From 1980 to 2010, clocks are 3,500x faster while architecture achieved perhaps a 50X improvement. “The silicon gave us way more than the architects could have made up for,” Colwell contends.
“So what happens when the energizer-bunny silicon engine runs out and we have to resort to being clever?” he asks. “How much runway is left? I’m saying we’re going to do the best we can, but don’t expect 3,500X.”
As for the argument that computer designers will continue to find other ways to make chips faster as long as there is a demand, Colwell contends that that will be true initially, but incremental tweaks cannot make up for the loss of an exponential.
Even though the DARPA rep is skeptical about a post-Moore’s law replacement, he points to several potential paths, some more promising than others:
- 3D stacking and improved packaging.
- Better cooling, longer battery life.
- Software, microarchitecture, architecture, I/O, memory (plenty of room for I/O and memory since still a first order impact on performance).
- Write new apps.
- Resilience (can be a difficult sell in general enterprise).
- Alternative switch technologies.
- DARPA perfect and upside, beyond convention.
- Better system interfaces (sensors, wireless, displays, human interface, etc.).
- Marketing.
While the end of free can be a scary thing, some of the biggest breakthroughs are the result of unyielding constraints. Look, for example, at the mobile computing industry, which has faced severe power and compute constraints. Now there is an entire ecosystem (i.e., ARM) built up around using “mobile” chips in servers.
If you haven’t watched this interesting and fast-paced keynote yet, check it out below: