Lesson 19 · Chips & LLMs

Transistor Physics: Why Scaling Got Hard

The device-physics floor beneath the entire industry — what a transistor is, why shrinking used to be free, why it stopped being free, and why that difficulty is itself the deepest moat.

Builds on: L4 (TSMC), L5 (packaging), L17 (EUV) Grounds: why only ~3 firms can play

You've studied the moats from the top down — TSMC's yield learning curve, ASML's EUV monopoly, advanced packaging. This lesson goes to the bottom: the single device every chip is built from, and the physics that turned "just make it smaller" from a free lunch into the hardest manufacturing problem on Earth. Understanding this tells you why the moats exist at all — and gives you a sharp BS-detector for node-name marketing.

Core thesis: For ~30 years, shrinking a transistor automatically made it faster, denser, and lower-power (Dennard scaling) — a free lunch that funded the whole industry. Around 2005 that lunch ended: voltage stopped scaling, leakage exploded, and progress required re-inventing the transistor's shape (planar → FinFET → GAA → CFET) and printing it with EUV. The escalating physics difficulty is exactly why the field of players collapsed from ~20 to ~3 — the device physics is the foundation of the TSMC/ASML moats.

01 — The MOSFET: A Voltage-Controlled Switch

The One Device Everything Is Made Of

A logic chip is billions of copies of one device: the MOSFET (metal-oxide-semiconductor field-effect transistor). It has three terminals that matter — source, drain, and gate — and it works as a switch:

MOSFET cross-section — the switch
p-type silicon substrate (body) Source Drain channel (forms when gate is ON) gate oxide (insulator) GATE V_gate controls the switch ↓ current flows source → drain only when channel exists
Apply enough voltage to the gate (above the threshold voltage, Vt) and it pulls a conducting channel into existence beneath the oxide, so current flows source→drain — the switch is "ON" (a logical 1). Below Vt, no channel, no current — "OFF" (a logical 0). The gate is separated from the channel by a thin insulating oxide, so it controls the channel by electric field alone, drawing ~no current itself.

Modern logic uses CMOS — complementary pairs of n-type and p-type MOSFETs — arranged so that in a steady state almost no current flows; power is consumed mainly when transistors switch. Hold that idea: it's the key to what broke.

02 — The Free Lunch: Dennard Scaling

Why Smaller Used to Mean Faster, Denser, and Cooler

In 1974 Robert Dennard (IBM, also the inventor of DRAM) described the bargain that would define the next three decades. Shrink every dimension of a transistor by a factor k, scale the voltage down too, and:

Dynamic power ≈ α · C · V² · f α = activity · C = capacitance · V = supply voltage · f = frequency. Note the V² term — lowering voltage was the magic lever.

Because power depends on the square of voltage, scaling V down each generation crushed the per-transistor power even as you added more transistors and raised frequency. This is the engine behind both Moore's Law (transistor count doubling ~every two years — really an economic observation about falling cost-per-transistor, not a law of physics) and the decades of "free" clock-speed gains.

03 — The Wall: Why Dennard Scaling Ended (~2005)

Voltage Stopped Scaling and Leakage Took Over

The free lunch ran on shrinking voltage. But voltage can't fall forever. To switch the transistor on, Vdd must stay comfortably above the threshold voltage Vt; and lowering Vt to match makes the "off" transistor leak. Two leakage currents blew up as dimensions shrank:

By ~2005, Vdd had stalled around ~1 V and couldn't go lower without leakage (static power) dominating. With the V² lever gone, power density stopped falling — the power wall. The visible symptom: clock speeds plateaued at ~3–4 GHz, where they remain ~20 years later. The industry's response was to stop chasing frequency and go parallel (multicore), then specialized (GPUs, then AI accelerators) — the entire premise of this course.

The pivot that created your portfolio: When single-thread scaling died, performance had to come from parallelism and specialization instead. That is why GPUs (L6), TPUs/ASICs (L9–L10), and advanced packaging (L5) became the growth story. The end of Dennard scaling is the root cause of the AI-hardware boom.

04 — Re-Inventing the Transistor's Shape

Planar → FinFET → GAA → CFET

With voltage scaling dead, the remaining problem was electrostatic control: as the channel got shorter, the flat (planar) gate — touching the channel on only one side — lost its grip, letting the drain leak current even when "off" (short-channel effects). The fix was to wrap the gate around more sides of the channel.

StructureEra / nodeGate wraps…Why it was needed
Planar MOSFETuntil ~22 nm1 side (top)Simple, but loses channel control as it shrinks → leakage
FinFET~22 nm (Intel 2011) → 5 nm3 sides (a vertical "fin")Standing the channel up as a fin lets the gate grip 3 sides → far better control, less leakage
GAA / nanosheet~3 nm / 2 nm (Samsung 2022, TSMC N2)4 sides (all around)Stacked horizontal sheets fully enclosed by gate → control at sub-3 nm dimensions
CFETresearch, ~end of decade4 sides, stacked n-over-pStack the complementary transistors vertically to keep area shrinking when 2D shrink is exhausted
The gate's grip on the channel — more sides, more control
Planar gate: 1 side FinFET gate: 3 sides GAA / nanosheet gate: 4 sides (stacked sheets) channel gate
Each structure surrounds the channel on more sides, restoring the electrostatic control lost to shrinking. This is why a "2 nm" GAA node is a genuinely harder manufacturing feat than the name suggests — and why EUV (L17) is required to pattern these structures at all.

Two more frontier tricks worth knowing: backside power delivery (Intel PowerVia, TSMC's A16) routes power on the back of the wafer, freeing the front for signal wiring and improving efficiency; and high-k metal gate (Intel, 2007) replaced the leaking SiO₂ oxide with hafnium-based material to fight gate tunneling.

05 — Why "nm" Is Marketing

Your Node-Name BS-Detector

Critical for diligence: the node name ("5 nm", "3 nm", "2 nm") no longer corresponds to any physical dimension on the chip. No feature in a "3 nm" process is 3 nm. The names became marketing labels around the 22 nm era. What actually matters:

Use this directly: when a company touts an "N nm" chip, ignore the number and ask for density and PPA versus the specific competing node. This is exactly the kind of claim financial media repeats uncritically — and where your technical taste (the mission) pays off.

06 — Investment Implications

Physics as the Foundation of the Moats

The device-physics floor under TSMC
Getting FinFET, then GAA, then CFET to work at yield is brutally hard and gets harder each node. That escalating difficulty is precisely why the field collapsed from ~20 leading-edge players to ~3 (TSMC, Samsung, Intel). The yield learning curve (L4) is the economic face of this physical difficulty — they're the same moat seen from two angles.
Slowing scaling raises the value of everything else
When raw shrink no longer delivers, differentiation moves to specialization (GPUs/ASICs, L6/L10), advanced packaging (chiplets/CoWoS, L5), and software (CUDA, L6). The end of Dennard scaling is the structural reason these became the value-capture layers — and why packaging/EDA/ASIC names matter more, not less, as Moore's Law slows.
Node transitions = competitive reset points
The FinFET→GAA shift is a rare moment where leadership can move. Samsung went GAA first (3 nm) hoping to leapfrog TSMC; Intel's 18A (RibbonFET + PowerVia) is its comeback bet. Watch GAA yield and flagship customer wins at each transition — that's when share actually changes (a TSMC falsifier from THESIS.md).
"nm" as a diligence filter
Evaluate nodes on density + PPA, never the marketing name. A company claiming node leadership on name alone is a yellow flag. This is the BS-detector that lets you read a process-technology announcement the way an engineer would, not the way a press release wants you to.
Primary Source

Go Deeper

Watch first: Asianometry — transistor-scaling and FinFET/GAA explainers — clear, engineer-level history of how each transistor structure solved the leakage problem. Read: the lithography & device chapters of Chip War (in RESOURCES.md) for how the physics drove the industry's consolidation, and Semiconductor Engineering for current GAA/CFET/backside-power coverage.

Comprehension Check

Quiz — 5 Questions

Select the best answer for each.

1. What does the gate voltage of a MOSFET actually control?

The amount of current the gate itself draws from the supply
Whether a conducting channel forms, allowing current source→drain
The clock frequency of the entire chip
The thickness of the gate oxide layer

2. Under Dennard scaling, why did shrinking transistors keep power density roughly constant?

Voltage scaled down too, and dynamic power depends on V²
More transistors always means less total power
Smaller transistors generate no heat at all
Clock frequency was held fixed every generation

3. Around 2005, Dennard scaling ended primarily because:

EUV machines were not yet available
Voltage could no longer scale down without leakage current exploding
Silicon ran out and had to be replaced with gallium
Customers stopped wanting faster chips

4. The progression planar → FinFET → GAA is fundamentally about:

Making the gate wrap the channel on more sides to restore electrostatic control
Switching from silicon to carbon nanotubes
Increasing the supply voltage each generation
Replacing transistors with optical switches

5. An investor reads "our new 2 nm chip." The technically correct response is:

2 nm means the gate length is exactly 2 nanometers
The name is marketing; ask for transistor density and PPA vs the specific competing node
All foundries' 2 nm nodes are equivalent by definition
Smaller node names always guarantee lower cost per chip
From your instructor: The chain to remember — transistor = voltage-controlled switch; Dennard made shrinking free; ~2005 voltage stopped scaling and leakage killed the free lunch; the fixes (FinFET→GAA→CFET + EUV) are so hard only ~3 firms remain. That difficulty is the physical bedrock of every moat in this course. Ask me anything: subthreshold leakage, why GAA beats FinFET, or how backside power delivery works.