The numbers Brief Episode map Takeaways Charts Quotes Clips Glossary Go next Transcript

Episode 25 · Deep tech · 52 min

Two Kinds of Computers

The engineer behind four decades of landmark silicon says the CPU-versus-GPU debate is over: soon there will be AI computers and deterministic computers, collaborating like cortices in one brain. Tenstorrent's bet — a graph-shaped data-flow chip and a clean-slate RISC-V team in India — aims at the 80% of world compute Keller says goes AI within ten years.

Deep tech & hardware AI & machine learning Leadership & org

Jim Keller

CTO & President, Tenstorrent · with Vishal Krishna

51:38

Said in this episode

▶ 41:02

80%

World compute that goes AI

Keller is 'very convinced' 80% of the world's computing will be AI in less than ten years.

▶ 17:12

10%/yr

The compound that becomes night-and-day

A 10% faster computer is unnoticeable in any one year but compounds over ten into twice the speed, half the size or twice the battery life.

▶ 41:50

30 → 4–5

Graphics startups after consolidation

At one point there were 30-odd graphics startups; the field reduced to four or five — the precedent Keller applies to today's ~50 AI chip companies.

▶ 35:28

130 → 7 nm

What lithography learned to write

Optical lithography at 130 nm got down to writing 7 nm transistors before the switch to 13.5 nm EUV light — with roughly a 100x density runway still on silicon.

▶ 29:25

10,000

Intel team size — 'a medium-sized city'

At Intel his team was about 10,000 people — a village needing roads and plumbing, versus the startup that aims like a bullet.

▶ 4:24

10×

His DEC timing verifier vs the incumbent

At DEC he and a colleague rewrote the timing verifier that sets chip frequency to be ten times faster and far simpler to use — learning architecture by doing.

The brief

The argument in sixty seconds

Keller's claim is that the CPU-GPU-accelerator taxonomy is collapsing into something sharper: AI computers, trained on data, and deterministic computers that run human-written code, collaborating the way a cerebral cortex works with a motor cortex. A neural network is computation organised as a graph, so Tenstorrent lays that graph across arrays of processing elements and lets data flow through — several times more efficient than a GPU, he argues, once the compiler makes new PyTorch models fast by default. Around the thesis sit his operating rules: Moore's Law is fine at TSMC; ten per cent a year compounds into night-and-day over a decade; build winning products, not novel papers; ten thousand people is a city and a startup is a bullet. He is convinced eighty per cent of computing goes AI within ten years — and Tenstorrent's new India office, designing a clean-slate RISC-V CPU, is hired for that bet.

Worth your time if you are

Semiconductor and systems engineers

Deep-tech founders making architecture bets

Engineering leaders scaling teams

Students plotting a first job in hardware

Episode map

Where the conversation travels

Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.

Deep tech & hardwareAI & machine learningLeadership & orgHiring & talentProduct strategyGCCs & services

01Cold open: from Harris to 14 years at DEC 0:00 Vishal introduces the fabless AI-chip company, then Keller retraces surfing into Harris, reading the VAX 11/780 manual en route to his DEC interview, 14 years of learning computer design by building tools, and why weekends exist. 02Computers are always a little too slow 6:12 Six decades of progress haven't changed the designer's experience — computation times transistors stays roughly constant because humans tolerate two-week jobs and dismiss sub-second ones. 03Two kinds of computers 8:18 Keller's thesis: AI computers trained on data and deterministic computers running human-written code will collaborate like cortices in one brain — and neural networks now model physics no equation describes. 04Why AIs will probably dream 12:09 Dreaming as the brain's nightly reorganisation — where Keller says almost all his good ideas arrive — suggests AI systems will get a dream mode that replays and restructures what they know. 05Moore's Law is fine; 10% a year compounds 14:55 Against the lock-it-for-seven-years business instinct, Keller says TSMC is still on the curve, consoles show five-year platforms work, 10% a year becomes night-and-day in a decade, and automotive is one iPhone moment from a scramble. 06Software 2.0 meets the data-flow machine 19:30 Karpathy's Software 2.0 becomes Tenstorrent's compiler problem: PyTorch models lowered to graphs laid across processor arrays, several times more efficient than CPUs or GPUs — once new models hit peak performance by default. 07India: from support teams to ownership 22:46 A decade of returning talent turned India's engineering offices into teams that own things; Tenstorrent has incorporated, taken its first ten accepts, and is handing them a clean-slate, verifiable RISC-V CPU. 08Run to win: products over papers 27:00 Papers select for novelty and products for winning — watch a hundred non-competing runners and you can only admire form, but in a race you know who won. 09Villages, bullets, and people problems 29:25 Intel's 10,000-person team is a medium-sized city that needs roads and plumbing, a startup aims like a bullet, and his father's law says technical problems get solved while people problems go on forever. 10EUV physics and the quantum reality check 34:22 Silicon keeps absorbing new atoms from the periodic table — 13.5 nm EUV light writing 7 nm transistors with a ~100x density runway — while photonics stays perpetually five years out and quantum has more media than use. 11The third hype cycle and the 50-chip race 38:30 GPT-3 is too expensive for a billion daily users and self-driving has gone sideways for five years, yet Keller is convinced 80% of computing goes AI within ten — with ~50 chip startups facing graphics-style consolidation, not winner-take-all. 12Basics, books, and Led Zeppelin row 25 44:13 Keller's advice — master basics, join a team that's going fast — gives way to Joscha Bach via Lex Fridman, Feyerabend's Against Method, the improbability of empire, and Led Zeppelin from the 25th row before a Bengaluru farewell.

Takeaways

Ideas to carry out of this hour

Two kinds of computers, and the line is sharp

The CPU/GPU/accelerator taxonomy is collapsing into AI computers — trained on data — and deterministic computers that run human-written C, Java or browser code. They will collaborate the way a cerebral cortex works with a motor cortex: an AI deciding to find a path or control something, then calling a deterministic program the way a thought instructs muscles. Keller insists the dividing line is sharp because the workloads, and the right circuits for them, are fundamentally different.

▶ 8:36 ▶ 9:20 ▶ 10:06

A neural network is a model without an equation

Classical science went data, then theory, then equation — E=mc², the index of refraction. A neural network goes straight from a large amount of data to a model that predicts what the system will do, with no equation in between. That makes very complicated, subtle phenomena tractable where a closed-form equation may not even exist — the basis of his claim that networks can understand physics better than the classical pipeline.

▶ 10:37 ▶ 11:22 ▶ 11:55

Technology is compound growth: 10% a year, ten years

The business instinct is to lock a chip platform for seven or eight years, and consoles really do hold for five — getting cheaper while games improve on fixed hardware. But technology developers take a small step every year: 10% faster is unnoticeable in any one year and night-and-day over ten — twice as fast, half the size, twice the battery life. Meanwhile TSMC keeps delivering whatever the Moore's-Law obituaries say, so the out-of-date platform eventually meets its iPhone moment, as automotive will.

▶ 15:53 ▶ 16:40 ▶ 17:12

Lay the graph on the chip: Tenstorrent's bet

An AI program is a large computation organised as a graph with data flowing through it, so Tenstorrent builds a data-flow computer that lays the graph across arrays of processing elements — several times more efficient than a CPU or GPU, he argues, and several times better than competition on price and efficiency. The harder half is the compiler: GPUs only hit peak on models that armies of engineers have hand-optimised with low-level libraries, while Tenstorrent wants a brand-new PyTorch model to reach good performance straight away. They have rewritten that compiler several times, learning each pass.

▶ 20:14 ▶ 21:32 ▶ 22:16

80% goes AI — through a graphics-style funnel

Keller is 'very convinced' 80% of the world's computing will be AI in under ten years, even while calling this the third AI hype cycle: GPT-3-class models are too expensive to serve a billion people daily, and ten autonomous-driving startups have gone sideways for five years. The endgame follows graphics — 30-odd startups consolidated to four or five — so today's roughly fifty AI chip companies face the same funnel. But it is not winner-take-all: applications are diverse, Google keeps the TPU in-house despite it being one of the best AI platforms on the planet, and how AI software should be written is still unsolved, which keeps the market open.

▶ 39:46 ▶ 41:02 ▶ 42:06

India's clean-slate RISC-V is a talent thesis

Keller traces Indian engineering's last decade: a huge return of experienced people, and local teams graduating from support work to genuinely owning things — he watched it happen inside AMD's and Intel's large India teams. Tenstorrent is opening an engineering office, deliberately not a sales office, with about ten accepts already in, and handing the team a from-scratch RISC-V CPU architected to be verifiable and modular using modern software techniques. His pitch to young engineers: at some point you need a clean piece of paper and original design you did yourself.

▶ 23:58 ▶ 25:32 ▶ 26:18

Move technical and people problems along together

His father's law: technical problems you solve, people problems go on forever. Frank Zappa preferred the synthesizer because it doesn't make mistakes — but you cannot build a large technical company with one person. The job is to meet people where they are without letting personal-problem work dominate the technical mandate; engineers who refuse people problems can't scale to a hundred, and managers who do nothing else never ship.

▶ 32:38 ▶ 33:08 ▶ 33:37

The numbers, drawn

What the episode measures

Every figure below was said on air — timestamps included, caveats kept.

Conversation share

portion of the hour spent on each theme

Deep tech & hardware32%

AI & machine learning26%

Leadership & org12%

Hiring & talent10%

Product strategy10%

GCCs & services6%

Computed from the chapter map of this episode.

Chip booms end in consolidation

companies

As stated in conversation — '30-odd' graphics startups fell to 'four or five'; ~50 network-processor startups to 'five to ten'; ~50 AI chip companies face the same funnel.▶ 41:50

The scales lithography learned to write

Keller's lithography arc as stated in conversation: 130 nm optics wrote 7 nm transistors via tricks ~10x smaller than the light itself; 13.5 nm EUV continues the shrink.▶ 35:12

Worth keeping

Lines that stay

Pretty soon there's going to be two kinds of computers: there's AI computers and deterministic computers.

— Jim Keller ▶ 8:36

If your computer was 10% faster you wouldn't notice. But 10% for 10 years could be night and day — twice as fast, half the size, twice the battery life. Technology is about compound growth.

— Jim Keller ▶ 17:12

There's two kinds of problems in the world: technical problems and people problems. Technical problems you solve, and people problems go on forever.

— Jim Keller, quoting his father ▶ 32:38

Quantum computing has been in the news for longer than it's been useful — it's gotten more media than use.

— Jim Keller ▶ 38:30

Clips that travel

Short on time? Start here

Deep-tech founders making architecture bets

Two kinds of computers, sharply divided

The episode's core thesis: AI vs deterministic machines, the brain analogy, and physics modelled without equations.

8:18 → 12:09 · 4 min ▶ Watch clip

Product leaders setting multi-year roadmaps

Moore's Law, consoles, and 10% compounded

Why platforms hold five years while technology compounds 10% a year — ending in the iPhone warning for carmakers.

14:55 → 19:30 · 5 min ▶ Watch clip

Students plotting a first job in hardware

Why Tenstorrent India gets a clean sheet

India's shift from support work to ownership, ten accepts in, and a from-scratch RISC-V CPU as the draw.

22:30 → 27:00 · 4 min ▶ Watch clip

Engineering leaders scaling teams

Products beat papers; villages vs bullets

Run to win the race, not for form — then Intel's 10,000-person city, the startup bullet, and the three kinds of people on teams.

27:00 → 32:24 · 5 min ▶ Watch clip

Investors sizing the AI compute market

The third hype cycle and the 80% claim

GPT-3 economics, self-driving going sideways, 80% of compute going AI, and why ~50 chip startups face a funnel.

38:30 → 44:13 · 6 min ▶ Watch clip

Glossary

The jargon, unpacked

RISC-V

An open-standard, licence-free instruction-set architecture — the base for the clean-slate CPU Tenstorrent's new India team is designing from scratch.

Data-flow computer

A chip where the program's graph is laid out across processing elements and data flows through the layout as the execution — Tenstorrent's architecture, versus fetching instructions sequentially on a CPU or GPU.

Software 2.0

Andrej Karpathy's term for programs created by training a network on data rather than hand-writing code (Software 1.0). Keller expects a 3.0 that nobody has defined yet.

EUV lithography

Chip printing with 13.5 nm extreme-ultraviolet light, succeeding the 130 nm optical era whose tricks were already writing 7 nm transistors — features ~10x smaller than the light used.

Fabless

A chip company that designs silicon but contracts the manufacturing to a foundry — Tenstorrent's model: design the AI chip and software, have a fab make the hardware.

TPU

Google's in-house Tensor Processing Unit — in Keller's words one of the best AI platforms on the planet, but never sold externally, which keeps the merchant AI-chip market open.

Connections

If this resonated, go here next

EP 01 · Deep tech

Full transcript

The whole conversation, searchable

Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.