AI Chip Process

AI Chip Process: Sub-2nm Gate-All-Around (GAA), Atomic Layer Processing, and Manufacturing Physics

An AI chip process is the physical and chemical sequence that turns raw monocrystalline silicon into a high-density, multi-billion-transistor accelerator. At the sub-2nm node regime, classical planar CMOS and FinFET architectures reach their electrostatic breakdown limits: gate length shrinkage causes source-to-drain quantum tunneling, severe drain-induced barrier lowering (DIBL), and uncontrollable subthreshold leakage.

To overcome these quantum limits and sustain the exponential compute density demanded by transformer models, foundries transitioned to Gate-All-Around (GAA) nanosheets paired with atomic-scale processing (Atomic Layer Deposition, Atomic Layer Etching, Extreme Ultraviolet Photolithography) and Backside Power Delivery Networks (BSPDN).


The Anatomy of a Sub-2nm GAAFET Transistor

Inside a Sub-2nm Gate-All-Around (GAA) Transistor atomic cross-section: nanosheet ribbons, high-k metal gate, and dual-sided interconnects Frontside Signal Wire (M1 Copper) Source Si:P / SiGe Drain Si:P / SiGe Metal Gate (TiN / W) Si Nanosheet 1 (30nm × 5nm) Si Nanosheet 2 (30nm × 5nm) Si Nanosheet 3 (30nm × 5nm) Bottom Dielectric Isolation (BDI - SiO₂) Ru BPR Nano-TSV → Backside VDD Power Mesh (Thick Cu Rail) High-k ALD Gate Dielectric HfO₂/ZrO₂, 360° wrap Atomic Nanosheet Channel Tinv<0.7nm · Ion>1.2mA/µm Buried Power Rail Ru, >950°C stable Gate-all-around wraps electrostatic control on all 4 sides of the channel — cutting subthreshold leakage below 100 pA/µm Decoupled Backside Power (BSPDN) routes supply current directly into source/drain without congesting frontside signal metal

The Fundamental Physics: Quantum Electrostatics and Atomic Processing

1. Gate-All-Around Electrostatic Scaling

In classical MOSFETs and FinFETs, the gate controls the channel across one, two, or three surfaces. As the physical channel length L_g scales below 12 nm, electrostatic field lines from the drain penetrate into the channel, causing severe Drain-Induced Barrier Lowering (DIBL) and gate control degradation.

The natural length lambda_g represents the spatial extent over which drain potential encroaches into the channel:

$$\lambda_g = \sqrt{\frac{\epsilon_{ch}}{2\epsilon_{ox}} t_{ch} \cdot t_{ox}}$$

Where epsilon_ch and epsilon_ox are channel and dielectric permittivities, t_ch is nanosheet channel thickness (t_ch ≈ 5 nm), and t_ox is equivalent oxide thickness (EOT ≈ 0.65 nm). To prevent punch-through and keep subthreshold swing near the theoretical thermodynamic limit (SS ≈ 60 mV/decade at 300 K), the physical gate length must satisfy:

$$L_g \geq 3 \cdot \lambda_g$$

By fully surrounding the silicon nanosheet channels on all four sides with high-kappa gate dielectric (HfO_2) and work-function metals, GAAFETs minimize lambda_g, allowing channel lengths to safely scale down to 10 nm to 12 nm with near-zero leakage.


Atomic Layer Etching & Selective Nanosheet Release

Atomic-Scale Channel Release: SiGe Selective Etching isotropic dry etching removes sacrificial SiGe without attacking atomic silicon ribbons 1. Epitaxial Superlattice Silicon Substrate Si₀.₇Ge₀.₃ (Sacrificial) Si Sheet 1 (Channel) Si₀.₇Ge₀.₃ (Sacrificial) Si Sheet 2 (Channel) Si₀.₇Ge₀.₃ (Sacrificial) Si Sheet 3 (Channel) Gas ALE 2. Selective SiGe Etch Silicon Substrate Si Sheet 1 (Intact) Si Sheet 2 (Intact) Si Sheet 3 (Intact) 150:1 Select 3. Suspended Ribbons Silicon Substrate Open 10nm Gap Open 10nm Gap Open 10nm Gap Suspended Si Ribbons Atomic Layer Etching achieves >150:1 selectivity: sacrificial SiGe layers are completely cleared without thinning the 5nm silicon channels This prepares the open vertical gaps for 360-degree conformal Atomic Layer Deposition (ALD) of HfO₂ gate oxide.
2. Atomic Layer Deposition (ALD) & Self-Limiting Surface Saturation

Once the sacrificial Si_0.7Ge_0.3 layers are cleared, the open 10 nm vertical gaps between suspended nanosheets must be lined with ultra-thin, pinhole-free dielectric and work-function metal layers. Standard Chemical Vapor Deposition (CVD) cannot coat high-aspect-ratio shadow regions uniformly.

Foundries deploy Atomic Layer Deposition (ALD), a sequential, self-limiting chemical reaction:
1. Pulse A: Precursor gas (e.g., hafnium tetrachloride HfCl_4 or TEMAH) is pulsed into the reactor, saturating surface hydroxyl groups (-OH) until steric hindrance halts adsorption at exactly one monolayer.
2. Purge A: High-purity nitrogen (N_2) or argon purges excess precursor molecules from the chamber.
3. Pulse B: Co-reactant gas (water vapor H_2O or ozone O_3) is introduced, reacting with the adsorbed hafnium complex to form a uniform HfO_2 monolayer while releasing HCl byproducts.
4. Purge B: Chamber is evacuated.

The cycle repeats N times, with film thickness t_film governed strictly by:

t_film = N · GPC

Where GPC is Growth-Per-Cycle (typically 0.08 - 0.12 nm/cycle). ALD delivers 100% step coverage across all four faces of the suspended nanosheet channel, maintaining sub-angstrom dielectric thickness uniformity across a 300 mm wafer.


Process Evolution: FinFET vs. GAAFET vs. Complementary FET (CFET)

Process Metric5nm / 3nm FinFET (N5/N3)2nm GAAFET Nanosheet (N2/A16)Sub-1nm CFET (3D Stacking)
Channel OrientationVertical 3D FinHorizontal NanosheetsMonolithic 3D Stacked n-over-p
Effective Gate Control3-Sided Gate4-Sided Gate-All-Around (360°)4-Sided Gate-All-Around
Effective Channel Width (W_eff)2 H_fin + W_fin ≈ 110 nm2 N (W_ns + T_ns) ≈ 210 nm2 N (W_ns + T_ns) in half footprint
Drive Current (I_on at V_DD=0.7V)~ 0.85 mA/µm~ 1.25 mA/µm (+45%)~ 1.30 mA/µm
Subthreshold Swing (SS)68 - 74 mV/dec62 - 65 mV/dec (near ideal)62 - 64 mV/dec
DIBL (Drain-Induced Barrier Lowering)45 - 65 mV/V<25 mV/V<20 mV/V
Power RoutingFrontside BEOL M0–M15Backside Power (BSPDN / BPR)Backside Power + 3D Interconnect
Standard Cell Area ShrinkBaseline15% - 20% reduction40% - 50% reduction

The Capital Compounding Flywheel and the Sub-2nm Process Moat

In semiconductor manufacturing, process leadership compounds through what Charlie Munger described as the cumulative volume learning curve:

C_N = C_1 · N^-b

Where C_N is the defect cost of the N-th wafer, C_1 is initial prototype cost, and b is the learning index (b ≈ 0.3 - 0.5).
* A foundry that processes 100,000 sub-2nm GAAFET wafers per month discovers and eliminates atomic-scale defect modes (such as nanosheet stiction, inner spacer dishing, and Ru BPR voiding) four times faster than a competitor processing 10,000 wafers.
* High yield (>85%) lowers die manufacturing costs, attracting marquee fabless AI designers (Apple, NVIDIA, AMD, Qualcomm).
* High cash generation funds the next generation of €380M High-NA EUV scanners and atomic layer cluster tools, widening the competitive barrier into an insurmountable capital toll bridge.

Take AI chip process further

Ask the copilot about this term, or have our engineers assess it against your process.