Engineering Philosophy: Jim Keller, Transistors Are Free

Key Takeaways

• Transistors are cheap; bottlenecks are expensive. Keller’s signature move, repeated across thirty years and a dozen chips, is to spend silicon to remove the thing standing in the way of throughput – wider pipelines, more execution units, faster interconnect – because a transistor you did not use is a transistor you wasted. The bet pays off when the extra hardware keeps the busy units fed instead of idle.¹⁷
• He has architected the chips behind half the computing you touch. DEC’s Alpha, AMD’s Athlon (K7) and K8 – where he co-authored the x86-64 (AMD64) instruction set and HyperTransport – then the Apple A4 and A5 behind the iPhone 4 and original iPad, then AMD’s Zen comeback, then Tesla’s self-driving computer, then a stint as an Intel SVP. Few engineers have shaped so many product lines.¹²³
• Moore’s Law is not dead – people just stopped counting the innovations under it. Keller’s public position, argued in talks and on the record, is that transistor scaling has a decade or two left because density comes from thousands of stacked innovations, not one trick. A modern transistor’s fin is still over a hundred atoms wide; there is a long way down.⁴⁷
• From a Penn State EE degree to RISC-V evangelist. Born around 1958, electrical engineering at Pennsylvania State University, then a career that ran DEC -> AMD -> Broadcom -> P.A. Semi -> Apple -> AMD -> Tesla -> Intel, and now CEO of Tenstorrent, building open AI accelerators on the open RISC-V instruction set.¹⁵⁶

The Principle

“While the world thinks Moore’s Law’s dead, the fabs and the technologists think it’s not, and everybody’s announced now a 10-year roadmap for Moore’s Law.” – Jim Keller⁴

Most engineers treat hardware as a fixed budget. You are handed a transistor count, a power envelope, a process node, and you optimize within it – shave a cycle here, fold a stage there, ration the silicon. Keller works the opposite way. He treats transistors as the cheap resource and the bottleneck as the expensive one. If a single unit is starving the rest of the machine, the answer is rarely to make that unit cleverer; it is to spend more silicon – duplicate it, pipeline it, widen the path to it – so the bottleneck disappears and everything downstream stays busy.⁷

The phrase the industry attaches to this is “transistors are free.” It is not literally true, and Keller knows it – silicon costs money and power. The point is comparative. Engineer time is scarce, design risk is dangerous, and an idle execution unit is pure waste. Against those, the marginal transistor is the thing you have the most of, and Moore’s Law keeps handing you more every couple of years. So the disciplined move is to spend the abundant resource to conserve the scarce ones. That single reframing – which resource is actually cheap – is why his chips tend to be wide, aggressive, and throughput-hungry rather than clever and narrow.¹⁷

It also explains why he is so loud about Moore’s Law not being dead. If transistors were about to stop getting cheaper, the whole strategy would collapse. So Keller does the unglamorous work of arguing, with numbers, that there is a decade or two of scaling left – that “Moore’s Law” was never one innovation but a cascade of thousands, each with its own diminishing-return curve, summing to an exponential.⁷ Spend the cheap resource to kill the bottleneck, and keep proving the cheap resource is still cheap. Everything else is detail.

Context

James B. Keller was born around 1958 and took a B.S. in electrical engineering from Pennsylvania State University, graduating in 1980.¹ What followed is one of the most improbable resumes in the history of the industry – not because of any single chip, but because of how many era-defining ones there are, across companies that are usually rivals.

He joined Digital Equipment Corporation in 1982 and stayed until 1998, working on the VAX 8800 and then the Alpha line – the 21164 and the out-of-order 21264 (EV6) – which were the fastest microprocessors of their day.¹ In 1998 he moved to AMD, where he helped launch the Athlon (K7) and was lead architect of the K8 microarchitecture. K8 is the one that mattered most: Keller co-authored the x86-64 (AMD64) instruction set that extended x86 to 64 bits, and the HyperTransport interconnect that tied multiple processors together. AMD64 became the 64-bit standard the entire PC and server world runs on, Intel included.¹²

Then the wandering began. SiByte in 1999 (MIPS networking chips), acquired by Broadcom in 2000 where he was chief architect through 2004; P.A. Semi from 2004 as VP of engineering, building low-power mobile processors. Apple acquired P.A. Semi in 2008, and Keller led the design of the A4 and A5 systems-on-chip – the silicon inside the iPhone 4, iPhone 4S, the original iPad, and iPad 2. Those chips seeded Apple’s in-house silicon program, the lineage that eventually produced Apple Silicon.¹³ He returned to AMD in 2012 to architect Zen, the microarchitecture that pulled AMD back from near-irrelevance into genuine competition.¹ Then Tesla from 2016, VP of Autopilot hardware, where he and Pete Bannon led the Full Self-Driving (FSD) computer; then Intel from 2018 to 2020 as senior vice president of silicon engineering.¹² Since late 2020 he has been at Tenstorrent – CTO, then CEO from 2023 – building AI accelerators on RISC-V.⁵⁶

The throughline is not loyalty to a company. It is loyalty to a method portable enough to win at DEC, AMD, Apple, Tesla, and Intel in turn.

The Work

Pipelining and “transistors are free”

Start with the idea that underwrites everything else, because it is where Keller’s instinct becomes arithmetic. A processor executes instructions in steps – fetch the instruction, decode it, read its operands, do the arithmetic, write the result. The naive machine runs one instruction all the way through those steps before starting the next. The problem is that while the arithmetic unit is working, the fetch and decode hardware sits idle; while the result is being written, almost everything else does nothing. Most of your expensive silicon is dark most of the time.

The fix is pipelining: split the work into stages and let them overlap, like an assembly line. While instruction one is in the arithmetic stage, instruction two is being decoded and instruction three is being fetched. Each instruction still takes the same number of steps, but the machine now finishes one every cycle instead of every five, because no stage is ever idle. The catch is that overlapping work costs hardware – latches between stages, logic to track dependencies, machinery to handle the cases where instruction two needs a result instruction one has not produced yet. You pay transistors to keep the units busy.¹

Keller’s chips push that bet hard. Going superscalar – multiple execution units so you finish more than one instruction per cycle – and out-of-order – reordering instructions so a stalled one does not block the ready ones – both cost a great deal of silicon for bookkeeping. The Alpha 21264 was an aggressive out-of-order design; AMD’s K7 and K8 were wide superscalar machines; Zen widened the path again.¹ In every case the reasoning is the same: the execution units are the point, idle units are waste, and the transistors needed to keep them fed are the cheapest thing in the building. Spend silicon to remove the bottleneck. That is the principle, and the pipeline is its simplest possible form.

AMD’s comeback and the AMD64 lineage

The work that reshaped the industry most was K8. By the early 2000s the 32-bit address space of x86 was running out of room – 4 GB of memory was becoming a real ceiling. There were two ways forward. Intel’s bet, Itanium, was to abandon x86 and build a clean new 64-bit architecture, breaking compatibility with the mountain of existing software. AMD’s bet, which Keller led the architecture of, was the opposite: extend x86 to 64 bits while keeping it able to run all the existing 32-bit code at full speed. That was x86-64, also called AMD64.¹²

The pragmatic bet won decisively. Software did not have to be rewritten; the upgrade path was painless; performance on legacy code did not suffer. AMD64 became the standard, and Intel eventually adopted AMD’s extensions rather than the reverse – the architecture in essentially every PC and server CPU today is the one Keller’s team specified. Alongside it, HyperTransport gave AMD’s Opteron servers a fast, point-to-point way to wire multiple processors together, attacking the memory-and-interconnect bottleneck that the “transistors are free” instinct always hunts for.¹² When Keller returned to AMD in 2012 to architect Zen, he repeated the pattern on a company that had nearly fallen out of the high-performance race: a wide, clean, modular core that closed the gap with Intel and gave AMD a credible product line again.¹ Twice he walked into AMD and twice he left it with the architecture that defined its next decade.

Apple silicon, from the ground up

The chapter with the longest shadow is the quietest. When Apple acquired P.A. Semi in 2008, Keller led the team that designed the A4 – Apple’s first in-house system-on-chip, shipped in the iPhone 4 and the original iPad in 2010 – and its successor the A5, in the iPhone 4S and iPad 2.¹³ Before this, Apple bought its mobile processors off the shelf. The A4 was the moment Apple decided to control its own silicon, and the team and discipline Keller helped establish became the foundation of the program that now produces the A-series and M-series chips powering every iPhone, iPad, and Mac.

The strategic logic is the same throughput-and-control instinct in a new domain. A phone’s hardest constraint is performance per watt: you cannot brute-force it with a desktop power budget. Owning the design end to end – rather than accepting a vendor’s general-purpose part – lets you spend your transistors exactly where the workload needs them and nowhere it does not. That is the “transistors are free” principle inverted for a battery: not “add silicon freely” but “place every transistor deliberately, because you own the whole design.” The line from the A4 to today’s Apple Silicon, and to the hardware-aware performance work John Carmack made famous, runs straight through that 2008 decision.³

First principles, and Tenstorrent’s open hardware

The method that ties the companies together is first-principles thinking – the willingness to throw out inherited assumptions and ask what the problem actually requires. Keller is blunt that architectures rot: roughly every five years, he argues, you should do a design from scratch rather than patch the old one, because the rewrite ends up both faster and less complicated than the accreted version it replaces.⁷ The discipline is to keep asking what you are really trying to do, stripped of the constraints that were true two process nodes ago and are not true now.⁷

That instinct is why he refuses the “Moore’s Law is dead” consensus. His argument is mechanical, not faith-based: transistor density is the sum of thousands of independent innovations, each on its own diminishing-return curve, and the sum is still exponential. A FinFET’s fin is over a hundred atoms wide today; you could imagine a transistor ten atoms on a side, “a million times smaller,” before you hit the floor. “So we’re not running out of atoms,” as he puts it.⁴⁷ The pessimists, in his telling, are counting one innovation and missing the cascade.

At Tenstorrent, where he is CEO, the first-principles bet is institutional. The company builds AI training and inference accelerators on RISC-V – the open instruction set, free of license fees and proprietary control – and intends to open-source its software stack and license its CPU and AI-core IP, not just sell chips.⁵⁶ Keller’s public conviction is that “in the next 5 to 10 years, RISC-V will take over all the data centers.”⁵ The wager is that open hardware, like open software before it, wins on first principles: lower the barriers, let many parties build, and the open ecosystem out-innovates the closed one. It is the same man who made x86 the standard now betting that the standard should be one nobody owns.

The Method

Read across the Alpha, AMD64, Zen, the A4, and Tenstorrent, and the same moves recur. Keller’s method is less a slogan than a set of standing commitments.

Spend the cheap resource to kill the bottleneck. The defining habit is to identify which resource is actually abundant – usually transistors – and spend it freely to remove whatever is starving throughput. Wider pipelines, more units, faster interconnect. The general lesson transfers: find the resource you have the most of and trade it against the one you have the least of, rather than rationing all of them equally. It is quality is the only variable at the silicon level – correctness and throughput are the goal, and the transistor budget is not the constraint you think it is.⁷

Rebuild from scratch on a schedule. Roughly every five years, do the design over rather than patch it, because inherited assumptions calcify and the rewrite comes out simpler. The courage to discard working-but-dated work is rare and load-bearing – the same instinct that lets Linus Torvalds throw out a subsystem that no longer fits.⁷

Ask what the problem actually requires, not what the last design assumed. First-principles thinking means stripping away the constraints that were true two nodes ago and are not true now. The discipline is to keep asking “what are we really trying to do” until the inherited baggage falls away – the evidence gate pointed at your own assumptions rather than at someone else’s claim.⁷

Place every transistor deliberately when the budget is tight. The flip side of “transistors are free” is the phone: when power, not area, is the wall, you own the whole design so you can put silicon exactly where the workload needs it. Knowing which regime you are in – abundant or scarce – and designing accordingly is the actual skill, not a blanket rule.³

Bet on openness as a first principle. RISC-V and an open software stack are a wager that lowering barriers lets more people build and the open ecosystem out-innovates the closed one. It is the minimum worthy product reasoning applied to a whole platform – ship the open thing that others can build on, rather than the closed thing only you can.⁵⁶

Influence Chain

Who Shaped Him

DEC and the Alpha culture. Keller learned high-performance design at Digital, on the Alpha line, in an organization that prized raw speed and aggressive out-of-order execution above almost everything. The conviction that you spend silicon to win on throughput was forged there, on the fastest microprocessors of their era. (Formative influence)

Dirk Meyer and Pete Bannon. Keller’s most important work came in partnerships – co-architecting Alpha’s out-of-order EV6 with Dirk Meyer, then leading Tesla’s FSD computer alongside Pete Bannon, who followed a parallel path into Apple’s silicon program. The collaborations are not incidental; the hardest chips are team sports, and Keller’s recurring co-architects shaped the work as much as he shaped theirs. (Direct influence)

Moore’s Law itself. The entire “transistors are free” strategy depends on transistors continuing to get cheaper. Keller’s worldview is downstream of Gordon Moore’s observation – which is why he defends it so fiercely. If the cascade of scaling innovations stops, the method has to change. (Formative influence)

Who He Shaped

Apple Silicon. The A4 and A5 team and discipline Keller helped stand up at Apple after the P.A. Semi acquisition became the foundation of the program that produces every A-series and M-series chip today. The most consequential consumer silicon of the last decade traces to a decision he was central to.

AMD’s two comebacks. K8/AMD64 made AMD a server contender and set the 64-bit standard the world runs on; Zen pulled AMD back into the high-performance race a decade later. Both architectures defined AMD’s competitive position for years after he left.

The RISC-V and open-hardware movement. As one of the industry’s most credible architects betting publicly and commercially on open instruction sets, Keller lends weight to the argument that hardware can follow software into openness. Tenstorrent is the demonstration.

The Throughline

Keller is where this series’ thread about spending the right resource meets the metal. John Carmack squeezed impossible performance out of fixed consumer hardware by understanding the machine down to the cycle; Keller works the layer below him – he designs the machine, and his answer to a bottleneck is not only to code around it but to add the silicon that removes it. Bjarne Stroustrup built C++ on the principle of zero-overhead abstraction, that you should not pay for what you do not use; Keller’s “transistors are free” is the hardware mirror – spend on what keeps the machine busy, waste nothing on what sits idle. And where Andrej Karpathy describes “Software 2.0,” programs compiled from data rather than written by hand, Keller is building the silicon that the new workload demands – AI accelerators on open RISC-V, designed from first principles for a problem the old general-purpose CPU was never shaped for. Carmack says master the hardware you are given; Stroustrup says do not pay for what you do not use; Keller says: when the hardware is the bottleneck, build a better one – and keep proving there are still atoms to spend. (Series bridge)

What I Take From This

The lesson I keep from Keller is to find out which resource is actually cheap, then spend it without guilt. It is easy to ration everything equally – to treat every constraint as binding and optimize timidly within all of them at once. Keller’s habit is to notice that one resource is abundant and the others are scarce, and to trade hard in the direction that abundance allows. Transistors are cheap; idle units and engineer-time are expensive; so spend transistors to kill the bottleneck. In my own work the abundant resource is rarely silicon – it is often compute, or a model’s tokens, or the ability to regenerate a draft cheaply. The move is the same: stop rationing the thing you have the most of, and aim it at the thing that is actually blocking you.

The second lesson is the willingness to start over on a schedule. Keller’s claim that you should redesign from scratch every few years – because the rewrite comes out simpler than the patched-up original – runs against every instinct to protect work you have already done. But he is right that assumptions calcify, that constraints which were once true quietly stop being true, and that the accreted design carries the weight of all of them. The discipline is to periodically ask what the problem actually requires now, with fresh eyes, and to have the nerve to throw out the old answer when it no longer fits. The thing I built last year was built for last year’s constraints. Keller’s career is a long argument that the brave move – the one that actually wins – is to build it again from first principles.

FAQ

What is Jim Keller’s engineering philosophy?

Spend the cheap resource to remove the bottleneck. Keller treats transistors as abundant – “transistors are free” – and engineer-time, design risk, and idle hardware as the expensive things, so his chips spend silicon freely on wider pipelines, more execution units, and faster interconnect to keep throughput high.¹⁷ Underneath that is first-principles thinking: roughly every five years, redesign from scratch rather than patch, because inherited assumptions calcify and the rewrite comes out simpler.⁷

What chips did Jim Keller design?

A remarkable range across rival companies: DEC’s Alpha processors (including the out-of-order 21264); AMD’s Athlon (K7) and K8, where he co-authored the x86-64/AMD64 instruction set and HyperTransport interconnect; the Apple A4 and A5 systems-on-chip behind the iPhone 4 and original iPad; AMD’s Zen architecture on his return in 2012; and Tesla’s Full Self-Driving computer. He also served as a senior vice president of silicon at Intel from 2018 to 2020.¹²³

Why does Jim Keller say Moore’s Law is not dead?

Because, in his account, transistor density is not one innovation but the sum of thousands of independent ones, each on its own diminishing-return curve, summing to an exponential – and that sum still has a decade or two left.⁷ He points out the fabs and technologists have published 10-year roadmaps, that a modern FinFET’s fin is still over a hundred atoms wide, and that you could imagine a transistor ten atoms on a side – a million times smaller. “So we’re not running out of atoms.”⁴⁷

What is Tenstorrent and what is Jim Keller doing there?

Tenstorrent is an AI-chip company where Keller has been CEO since 2023 (CTO before that). It builds AI training and inference accelerators on the open RISC-V instruction set, intends to open-source its software stack, and licenses its CPU and AI-core IP alongside selling chips.⁵⁶ Keller’s bet is that open hardware will follow open software’s path to dominance – he has said he believes “in the next 5 to 10 years, RISC-V will take over all the data centers.”⁵

Sources