SemiconductorsProcess NodesTransistor PhysicsAI Chip Design

What "3nm" Actually Means: Process Nodes, Transistor Physics, and Why Scaling Is Slowing

When TSMC announces its N3 process or Intel announces 18A, what is actually changing about the transistor? What does the node number mean — and why does it no longer mean what it used to? This essay explains the physics of transistor scaling from FinFET to Gate-All-Around, why Moore's Law is slowing, and what it means for AI chip design.

By Manish KL · April 2026 · ~17 min read · Hardware Essay
Abstract. "3nm" is a marketing label, not a physical measurement. The gate length of the transistors in TSMC's N3 process is not 3 nanometres — it is approximately 12 nm. This essay explains what the process node number actually describes, how transistors work at the physical level (and why they stop working below certain dimensions), what FinFET and Gate-All-Around (GAA) architectures do to extend scaling, how TSMC N3/N2, Samsung 3GAE/2nm, and Intel 18A compare, and what the slowdown of Moore's Law means specifically for AI chip design — where more transistors translates directly into more tensor cores, more SRAM, and more arithmetic per watt.
~12 nmactual gate length in TSMC "N3" process — the "3nm" label is a marketing node name, not a dimension
~2×transistor density improvement per node — was 2× every 2 years (Moore's Law); now 2× every 3–4 years
$20B+cost to build a leading-edge fab (N3 or better) — up from $1B for a 130nm fab in 2000
~800 mm²practical maximum reticle-limited die size — the physics ceiling before chiplets become necessary
Contents
  1. What a process node number actually means
  2. How a transistor works — and why size matters
  3. Planar transistors and why they stopped scaling
  4. FinFET: the architecture that bought a decade of scaling
  5. Gate-All-Around: the transistor structure at N2 and beyond
  6. Transistor density: what each node actually delivers
  7. TSMC vs. Samsung vs. Intel: where the leading nodes stand in 2026
  8. What a node shrink actually buys for AI chips
  9. Why scaling is slowing — and what the alternatives are
  10. What this means for AI chip design through 2030

1. What a process node number actually means

The "3nm" in TSMC N3 does not refer to any single physical dimension of the transistor. It is a marketing label that originated in the 1990s when it did correspond to the transistor gate length — the physical dimension of the gate that switches the transistor on and off. At that time, a "250nm process" genuinely had ~250nm gate lengths. Shrinking the gate length was the primary way to pack more transistors per unit area.

Around the 28nm–16nm transition (2011–2014), the correspondence between node name and physical gate length broke down. The competitive dynamics of the semiconductor industry incentivised foundries to use smaller node numbers as marketing claims, even as the actual physical improvements became more multidimensional — better pitch density, new transistor architectures, tighter metal spacing — rather than simple gate length reduction.

Today, "3nm" describes a process generation — a collection of manufacturing improvements that together produce higher transistor density, better power efficiency, and higher performance compared to the previous generation. The actual gate length is approximately 12 nm. The relevant physical measurements are:

Physical metricWhat it measuresWhy it matters for AI chips
Gate pitchDistance between adjacent transistor gates (centre-to-centre)Determines transistor density along the gate dimension
Metal pitch (M0/M1)Spacing of the finest wiring layer connecting transistorsDetermines routing density and ultimately SRAM cell size
Cell heightHeight of a standard logic cell (in track units)Determines logic density — how many gates per mm²
SRAM bit cell areaArea of one 6-transistor SRAM bit cell in µm²Determines how much on-chip cache fits per mm² of die area
MTr/mm² (transistors/mm²)Total transistors per square millimetreThe headline density metric — how many logic gates per die

2. How a transistor works — and why size matters

A MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor) is a voltage-controlled switch. It has three terminals: the source (where current enters), the drain (where current exits), and the gate (which controls whether current flows). Between the source and drain is a thin channel of semiconductor material. The gate is separated from the channel by a thin insulating oxide layer.

When a sufficient voltage is applied to the gate, an electric field penetrates through the oxide and induces a conducting channel between source and drain — the transistor turns on. When the gate voltage is removed, the channel disappears and no current flows — the transistor turns off. This binary switching is the basis of all digital logic.

The gate length (L) is the distance between source and drain along the channel. Reducing L has three effects:

1
Higher switching speed. Electrons travel from source to drain in time proportional to L/velocity. Shorter L means faster switching. Clock speed and logic delay improve.
2
Lower switching energy. The capacitance of the gate (which must be charged and discharged each switching cycle) is proportional to gate area ≈ L × W. Shorter L means lower capacitance, lower charge per cycle, lower switching energy. This is the power efficiency gain from scaling.
3
Higher transistor density. Smaller transistors pack more tightly. More transistors per mm² means more logic gates, more SRAM, more tensor cores in the same die area.

These three benefits — speed, power, density — are the classic returns of transistor scaling, and they drove the semiconductor industry for fifty years. The problem is that below approximately 5–7nm gate length, physics starts fighting back.

3. Planar transistors and why they stopped scaling

Until approximately 22nm (Intel, 2011), all commercial transistors used a planar architecture — the channel is a flat region at the surface of the silicon substrate, and the gate sits on top of it, separated by a thin oxide. Scaling meant making the gate shorter and the oxide thinner.

Two fundamental problems emerged as gate lengths dropped below ~20nm:

1
Short-channel effects. When the gate length is very short, the electric field from the source and drain "leaks" into the channel region, competing with the gate's ability to control it. This makes it harder to turn the transistor fully off — current leaks even when the gate voltage is zero. This is called subthreshold leakage, and it increases power consumption even when the transistor is nominally "off."
2
Gate oxide tunnelling. The oxide separating the gate from the channel must become thinner as the transistor scales (to maintain electrostatic control of the channel). Below ~1.2nm oxide thickness, electrons quantum-mechanically tunnel through the oxide from gate to channel, creating gate leakage current even when the transistor is supposed to be off. This is not a manufacturing defect — it is quantum mechanics and cannot be engineered away in a planar geometry.

Both effects cause the same outcome: power consumption that cannot be reduced by lowering voltage, because the transistor leaks regardless of operating conditions. Planar transistors hit a physical wall at approximately 22nm gate length. A new architecture was needed.

4. FinFET: the architecture that bought a decade of scaling

The FinFET (Fin Field-Effect Transistor) solved the planar transistor's short-channel problem by changing the geometry of the channel. Instead of a flat channel at the silicon surface, the FinFET uses a thin vertical "fin" of silicon that protrudes from the substrate. The gate wraps around three sides of this fin.

Transistor architecture evolution: Planar → FinFET → Gate-All-Around Planar (≥22nm) channel oxide gate S D gate controls 1 side leakage below ~22nm FinFET (7nm–16nm) fin gate (3 sides) gate wraps 3 sides much better channel control GAA / Nanosheet (≤3nm) gate (all 4 sides) maximum gate control stacked nanosheets
Fig 1. Transistor architecture evolution. Planar transistors control the channel from one side — susceptible to short-channel effects below ~22nm. FinFET wraps the gate around three sides of a vertical silicon fin, dramatically improving electrostatic control. Gate-All-Around (GAA) / nanosheet transistors wrap the gate around all four sides of stacked horizontal silicon nanosheets — the maximum possible gate control, enabling continued scaling below 5nm.

By wrapping around three sides of the fin instead of sitting on top of a flat channel, the FinFET gate has dramatically more electrostatic control over the channel. This suppresses short-channel effects, allowing scaling to continue to approximately 5nm gate pitch. FinFETs enabled TSMC 16nm (2014) through TSMC 5nm (2020) and TSMC 4nm (2022), covering nearly a decade of Moore's Law continuation.

5. Gate-All-Around: the transistor structure at N2 and beyond

FinFETs have their own scaling limit. As the fin becomes narrower (to reduce footprint), quantum confinement effects change the electrical properties of the channel. Below approximately 5nm fin width, the FinFET's performance degrades. A new geometry is needed.

Gate-All-Around (GAA), also called nanosheet or nanowire transistors, wraps the gate around all four sides of the channel — not three. The channel is a horizontal sheet (or stack of sheets) of silicon, and the gate material surrounds it completely. This achieves the maximum possible electrostatic gate control, enabling continued scaling to sub-5nm pitch while maintaining acceptable leakage.

The specific implementation used by TSMC at N2 and Samsung at 3GAE is stacked nanosheets — multiple horizontal silicon nanosheets stacked vertically, with the gate surrounding each one. This not only improves electrostatic control but also increases drive current per unit footprint (by providing more channel area per device footprint), which is critical for performance at low supply voltages.

Intel's RibbonFET (in 18A and 20A) is their name for the same concept — a GAA device with stacked ribbon-like channels. The geometry differs slightly between manufacturers but the physics principle is identical: maximum gate coverage, maximum channel control.

6. Transistor density: what each node actually delivers

Process nodeFoundryYear (HVM)Logic density (MTr/mm²)SRAM cell (µm²)Transistor typeKey AI chips
7nm (N7)TSMC2018~910.027FinFETA100 (TSMC 7nm)
5nm (N5)TSMC2020~1710.021FinFETApple M1, various AI chips
4nm (N4)TSMC2022~1980.019FinFET (improved)H100 (TSMC 4N custom)
3nm (N3E)TSMC2023~2920.0171FinFET (3rd gen)Apple M3, A17 Pro
3nm (N3P)TSMC2024~320~0.016FinFET (optimised)H200, future AI GPUs
2nm (N2)TSMC2025–26~380+~0.014GAA nanosheetGB300 successors, Apple A19
3GAESamsung2023~145~0.020GAA nanosheetSamsung Exynos
18AIntel2025~200+~0.017GAA RibbonFETIntel GPU successors

A critical observation from this table: Samsung's 3GAE, despite being the world's first commercial GAA transistor (a genuine architectural advance), delivers lower density than TSMC's N3E FinFET. Transistor architecture is necessary but not sufficient — manufacturing maturity, defect density, and process optimisation determine actual density and yield. TSMC's FinFET at N3 outperforms Samsung's first-generation GAA at 3GAE because TSMC has spent years optimising its FinFET process while Samsung was first-generation on a new architecture.

7. TSMC vs. Samsung vs. Intel: where the leading nodes stand in 2026

The leading-edge semiconductor manufacturing market has consolidated to three credible players for nodes below 7nm: TSMC, Samsung, and Intel Foundry. Their competitive positions as of April 2026:

T
TSMC is the undisputed manufacturing leader. N3P is in high-volume manufacturing with strong yield and broad customer adoption (Apple, NVIDIA, AMD, Qualcomm, Broadcom). N2 (GAA nanosheet) is in risk production with HVM entry late 2025 / volume 2026. N2P (optimised N2) targets 2026–2027. TSMC's CoWoS packaging technology for 2.5D integration of GPU + HBM is also technically mature and running in high volume for H100/H200 production. TSMC's NVIDIA-specific "4N" custom process (between N4 and N3) is what the H100 uses — a customer-specific optimisation of N4 for GPU workloads.
S
Samsung Foundry was first to market with GAA (3GAE in 2023) but has faced yield challenges. 3nm yield rates have been reported as significantly below TSMC's comparable node. Samsung is investing heavily in process improvement and is a credible second source for customers seeking supply diversification. Their 2nm GAA node targets 2025–2026. Samsung also supplies HBM3e to NVIDIA (SK Hynix is the primary supplier; Samsung is secondary).
I
Intel Foundry is attempting a remarkable re-entry into leading-edge manufacturing after falling behind during the 10nm/7nm era. Intel 18A (GAA RibbonFET + PowerVia backside power delivery) is Intel's most ambitious node and targets density comparable to TSMC N2. Risk production in 2024, limited HVM 2025. Microsoft's Azure custom AI chip is an early 18A customer. Intel's integrated IDM (Integrated Device Manufacturer) model — making both chips and selling foundry capacity — creates unique dynamics that TSMC doesn't face.

8. What a node shrink actually buys for AI chips

For AI chip designers, a node shrink translates into four concrete benefits — and the magnitude of each has changed as scaling has slowed.

BenefitMechanismN7→N4 (historical)N4→N2 (current gen)AI chip implication
More transistors per mm²Smaller devices, tighter pitch~2.2× density~1.9× densityMore tensor cores, more SRAM per die area
Lower power per operationLower capacitance per transistor~25–30% reduction~15–20% reductionMore tokens per watt; lower cooling requirements
Higher frequency at same powerFaster transistor switching~10–15% gain~5–10% gainModest clock frequency improvement
More die area for same costSmaller die for same functionSignificantDiminishingCost reduction per unit compute slower than historically

The key trend: the gains per node are real but diminishing. N7 to N4 delivered approximately 2.2× transistor density, ~30% power reduction, and ~15% frequency gain. N4 to N2 is projected to deliver ~1.9× density, ~20% power reduction, and ~8% frequency gain. The returns are still positive — every generation of AI chip benefits from the node advance — but the magnitude of improvement per generation is smaller than it was in the FinFET scaling era.

Translating density improvement to AI chip performance — H100 → N2 equivalent
H100 (TSMC 4N custom): ~80B transistors in ~814 mm²
Equivalent chip at N2 (1.9× density): same area → ~152B transistors
Or: same transistor count → ~428 mm² die (47% smaller)

If transistors → tensor cores at same ratio:
H100 tensor cores: 528 (in 132 SMs)
N2 equivalent: ~1,003 tensor cores → ~1,877 TFLOP/s FP8 (vs. 989 today)

Or equivalently: same performance in half the die area → halved wafer cost per unit performance

9. Why scaling is slowing — and what the alternatives are

Moore's Law — the observation that transistor density doubles approximately every two years — has not ended, but it has slowed materially. The two-year cadence has stretched to three to four years for density doublings, and the cost per transistor has stopped falling (it actually increased at some nodes).

The causes are physical and economic:

1
EUV lithography cost. Below 7nm, features cannot be patterned with conventional deep ultraviolet (DUV) light — the wavelength is too long. Extreme Ultraviolet (EUV) lithography uses 13.5nm wavelength light (vs. 193nm for DUV), produced by firing a laser at a tin droplet to create plasma that emits EUV. A single ASML EUV scanner costs ~$200M and requires a factory of support equipment. Each lithography step costs more; more steps are required per node.
2
Multi-patterning complexity. Even with EUV, the finest features (below ~13nm pitch) require multiple exposures — patterning the same layer two or more times with different masks and alignment steps. Each multi-patterning step adds cost, time, and potential alignment error, reducing yield.
3
Quantum effects at the channel level. As nanosheets become thinner (approaching 5nm thickness), variability in individual atom placement becomes significant. A single misplaced dopant atom can change the transistor's threshold voltage measurably. Controlling this variability requires increasingly precise process control and wider design margins.

The industry's response has been to invest in packaging innovation as an alternative to planar scaling. Instead of making transistors smaller, connect more chiplets together with shorter, denser interconnects. HBM, CoWoS, EMIB, TSMC SoIC, and Intel's 3D Foveros are all variants of this strategy — achieving system-level performance improvements through better integration rather than better individual transistors.

10. What this means for AI chip design through 2030

For AI chip designers, the slowdown of transistor scaling creates a specific set of constraints and opportunities through 2030:

1
Die size becomes more important. If density improvement per node is ~1.9× rather than 2.5×, getting more transistors means using a larger die — up to the reticle limit of ~858 mm² (TSMC) or ~800 mm² (Samsung). The H100 at 814 mm² is already near this limit. Future AI GPUs will hit the reticle ceiling and be forced into chiplet designs (multiple dies in the same package) to add more compute.
2
SRAM density improvement is the critical metric. For inference specifically, the most valuable use of additional transistors is more on-chip SRAM (to hold weights closer to compute and reduce HBM trips). SRAM cell area at N2 (~0.014 µm²) vs. N4 (~0.019 µm²) is a 36% improvement — meaningful but not transformative. The SRAM density constraint is the reason wafer-scale approaches (Cerebras) remain architecturally attractive even as standard scaling continues.
3
Power efficiency gains matter more than raw density. A 20% reduction in power per operation at N2 vs. N4 directly reduces joules-per-token and cooling requirements. At 120–140 kW per rack, cooling is a hard constraint. A chip that delivers the same throughput at 20% lower power enables denser rack packing and lower operating cost — often more valuable than a density increase that requires more power.
4
The competitive advantage of leading-edge nodes is compressing. When TSMC N3 delivers 2.2× the density of N7, the chip designer on N3 has a structural advantage over the N7 designer. When N2 delivers 1.9× N4, the advantage shrinks. Combined with the 3–4 year cadence, this means AI chip roadmaps must increasingly look to architecture innovation (better tensor core designs, smarter memory hierarchies, more efficient data movement) rather than relying on node transitions to deliver performance improvements.
The "free lunch" of Moore's Law — where every two years you got twice as much compute for the same cost — is over for AI chip design. The next generation of performance improvement requires architectural creativity: better packaging (chiplets, advanced 3D integration), smarter on-chip memory management, and workload-specific silicon design. The transistor is still improving. It is improving more slowly, more expensively, and with smaller returns per generation. That changes the calculus for every AI chip programme launched after 2024.

What this essay does not cover
This essay focuses on logic transistors and their scaling. It does not cover: DRAM process nodes (which follow a different scaling trajectory from logic), the specific EUV and High-NA EUV equipment roadmap from ASML, backside power delivery (PowerVia / TSMC equivalent) as an additional scaling enabler, 3D transistor stacking (CFET — Complementary FET — which may extend scaling beyond 2nm), or the geopolitical dimensions of process node access (export controls, ASML equipment restrictions to China). Each of those is a substantial topic that deserves its own treatment.