The Diminishing-Returns Trap That Keeps Moore's Law Alive
Legendary microprocessor engineer Jim Keller recently sat down for a wide-ranging conversation on the AI Podcast. Keller's resume includes stints at AMD, Apple, Tesla, and now Intel, with key roles on the AMD K7, K8, K12, and Zen microarchitectures, the Apple A4 and A5 processors, and co-authoring the x86-64 instruction set and HyperTransport interconnect. In the interview, he offered a contrarian take on the perennial claims that Moore's Law is dead: the premise is wrong because the framing is wrong.
A Matter of Perspective
Keller's argument starts with the concept of a diminishing returns curve. A project improves, then plateaus. To reach the next level, you must start a new project whose initial output is lower than the old project's ceiling, but whose trajectory ends higher. This creates two anxieties: short-term disaster (fear of changing what works today) and long-term disaster (fear of being stuck on a plateau). Quarterly objectives, Keller says, make people terrified of change, while long-term builders see short-term limitations as the barrier to future success. You can run multiple projects concurrently, but not everyone will be happy.
The common misconception is that Moore's Law is a single curve. In reality, it is the aggregate of thousands of separate innovations, each with its own diminishing-return lifecycle. The combined result has been steady exponential improvement. When an expert on one particular curve sees its plateau, they declare the whole thing done—while other teams are pursuing entirely different approaches. That, Keller suggests, is just how the industry works.
Room Left on the Table
By his estimation, there is still substantial runway. A modern transistor measures roughly 1000 x 1000 x 1000 atoms. Quantum effects only emerge at the scale of 2-10 atoms, meaning a transistor could theoretically shrink to 10 x 10 x 10 atoms—a million-fold reduction in volume. Techniques already exist for depositing and even placing individual atoms. The practical constraint is throughput: if placing one atom takes ten minutes, assembling a computer's worth of material—on the order of 10^23 atoms—would be impractical. The innovation stack, in other words, remains deep and broad.
Keller expects the near term to deliver more transistors every two to three years, at a scale significant enough that computer architecture itself will need to change to exploit it.
Architecture Under Pressure
That change won't necessarily make things simpler. Keller dismisses the notion that computers should be simple and clean, noting the market for simple, clean, and slow computers is zero. Instead, he advocates for periodically discarding accumulated complexity: to make real progress in architecture, you should start from scratch roughly every five years.
The full interview covers far more ground than chip scaling. For anyone interested in how microprocessors are conceived and built, it's worth a listen.
