Jim Keller — Legendary microprocessor engineer who has worked at AMD, Apple, Tesla, and Intel. Known for AMD's K7/K8/Zen architectures, Apple's A4/A5 chips, and co-authoring the x86-64 instruction set specification.
The gist
Jim Keller walks through computer architecture from atoms and transistors up to instruction sets, branch prediction, and out-of-order execution. He argues Moore's Law will keep going another 10-20 years because it is really thousands of stacked innovations, each on its own diminishing-return curve. He distinguishes 'found parallelism' (CPUs) from 'given parallelism' (GPUs), and explains why deep understanding beats executing recipes. The conversation ranges into AI, autonomous driving (where he and Lex disagree on how hard the human-behavior element is), working with Elon Musk, first-principles thinking, and whether the universe is a computer.
Big reveals
- Keller says a good computer architecture should be rewritten from scratch every five years, far more often than the industry's typical 10+ years.
- He recounts being told Moore's Law would die in 10-15 years for his entire 40-year career, and decided to stop worrying about that prediction.
- Modern branch predictors use something resembling a neural network with tens of megabits to reach ~99% accuracy, up from 1000 bits for 85%.
- He asked his team for a roadmap to a 100x transistor shrink; they only got to 50x in two weeks, and he believes the full 100x is achievable.
- Keller claims 'you don't have to be especially smart to drive a car,' framing autonomy as mostly an attention problem computers excel at.
- He dismisses AI existential-threat fears, arguing a superintelligence would have interests astronomically different from fighting humans over dirt.
- He suggests if the universe is a computer it's a bizarre one, since simulating quantum effects in a tiny region seems to take near-infinite computation.
Things worth remembering
- About 90% of program execution runs on just 25 instructions/opcodes, which have been stable for 25 years.
- Modern CPUs fetch ~500 instructions at once, compute the dependency graph, and execute deeply out of order.
- To keep a 600-instruction window effective, processors must predict roughly 99 of 100 branches correctly.
- A modern transistor is about 1000x1000x1000 atoms; quantum effects appear around 2-10 atoms, leaving room to shrink ~1,000,000x.
- Keller reframes organizations like computer architectures, treating people as differing functional units.
- His key management insight: most people don't think simply enough, confusing following recipes with deep understanding.
- He says he's been awake reading ~50-55 years, and a good book compresses 20 years of someone's passionate work into 200 pages.
- At one company he read 19 more management books than any other VP, and half the techniques worked the first time.
- Two hard constants in chip design: people don't get much smarter, and teams can't grow much beyond ~100 before needing org boundaries.