01Executive summary
Three milestones that made computation a component rather than a facility, gave it persistent memory, and kept both getting cheaper for five decades.
The microprocessor of 1971 put a complete processor on one chip, so computation could be embedded in anything. Flash memory from around 1980 gave that computation non-volatile storage with no moving parts. Extreme ultraviolet lithography, in volume production from 2019 after roughly two decades of development, continued the feature scaling on which the economics of both depend.
02The microprocessor: generality as an economic strategy
The design that became the Intel 4004 originated in a calculator project. The customer wanted a family of custom chips, one per model. Ted Hoff proposed instead a single general-purpose processor whose behaviour would be set by a stored program, with Federico Faggin leading the silicon design and Masatoshi Shima representing the customer’s requirements.
The engineering argument is worth stating plainly, because it is a general one. A custom chip per product means the design cost is borne by that product alone. A general processor means one design cost is amortised across every product that uses it, and differentiation moves into software, where changes cost almost nothing to reproduce. Generality was not chosen for elegance; it was chosen because it moved cost from a fixed non-recurring charge to a nearly free variable one.
Priority for the microprocessor is genuinely contested. The 4004 was the first commercially marketed single-chip processor, announced in 1971. Texas Instruments developed a comparable calculator-oriented device in the same period. A single-chip processor developed by Four-Phase Systems and a processor set built for the F-14 aircraft’s central air data computer both predate or parallel it, and were disclosed later. The defensible statement is that several groups converged on the same architecture as soon as transistor density allowed, and that the 4004 is notable for being the first offered as a general-purpose product to anyone who wanted one.
Embedded computing became the default
Once a processor cost a few dollars, control logic that had been mechanical, hydraulic or hard-wired became software. Most microprocessors ever made sit in things nobody calls a computer.
The instruction set became an interface
Software written to an instruction set survives the hardware it was written for. This is the same interface-standardisation move as the shipping container, applied to computation, and it is why decades-old binaries still run.
03Flash memory: the storage hierarchy rearranged
Computer memory before flash divided cleanly. Volatile semiconductor memory was fast and lost everything at power-off. Magnetic media retained data without power but were slow, mechanical and fragile. Every system design worked around that gap.
Flash stores charge on a gate isolated by insulator, so the state persists without power. Its distinguishing constraint is that erasure happens in blocks rather than per byte, which is where the name comes from and which shapes everything about how it is used.
Flash succeeded despite constraints that look disqualifying on paper: it cannot be overwritten in place, it wears out, and early devices were more expensive per bit than the media they competed with. It won because the controller hid those constraints behind a conventional interface, and because the properties it did offer — no moving parts, low power, shock tolerance, small size — mattered more in the applications that grew fastest. Managing an awkward physical constraint in an abstraction layer, rather than solving it, is one of the most productive moves available in engineering.
04Lithography: the machine behind the trend
Semiconductor scaling is often discussed as though it were a natural law. It is not. It is the aggregate outcome of continuous, expensive improvement in photolithography, and every step of that improvement has been fought for against optical physics.
The resolution of a projection system improves with shorter wavelength and higher numerical aperture. For decades the industry advanced by shortening wavelength through the ultraviolet, then by raising numerical aperture, then — when both stalled at 193 nm — by placing water between the lens and the wafer to raise the effective aperture, and by printing a single feature with multiple patterned exposures.
- Shorter wavelengthSuccessive light sources through the ultraviolet, each requiring new optics, resists and masks.
- Higher numerical apertureLarger, more precise optics gathering light over a wider angle, with tighter depth-of-focus tolerance as a penalty.
- ImmersionWater between final lens and wafer raises the effective aperture, extending 193 nm well past its expected limit.
- Multiple patterningOne layer printed in several exposures, multiplying cost and overlay accuracy demands to reach smaller pitch.
- Extreme ultraviolet13.5 nm light, produced by a laser-driven tin plasma, reflected by multilayer mirrors in vacuum because it is absorbed by everything including glass and air.
At 13.5 nm there are no lenses, because every material absorbs. The optics must be reflective multilayer mirrors, each losing part of the light, in a vacuum. There is no conventional lamp, so the source is a droplet of tin struck by a high-power laser tens of thousands of times per second. Debris from that plasma must not reach the mirrors. The masks are reflective rather than transmissive and must be defect-free at a scale where a single atom-scale flaw prints. None of these is a scientific unknown; all of them are engineering problems that took two decades and extraordinary sustained investment. It is the clearest modern illustration of the pattern this series keeps returning to: the physics was never the hard part.
05Takeaways for current practice
- Generality is an economic argument, not an aesthetic one. It amortises fixed cost across applications and shifts differentiation to where reproduction is free.
- Hide awkward physical constraints behind an abstraction. Flash won by managing block erasure and wear in a controller rather than by eliminating them.
- Know the amplification factor in any indirection layer. Write amplification has direct analogues in caching, batching and retry logic, and it determines real-world cost.
- Treat exponential trends as the outcome of sustained investment. They are maintained deliberately and can stall; nothing about them is automatic.
- Expect the last increment of a mature technology to cost more than the first. Immersion and multiple patterning bought years at rapidly rising complexity.