Hardware Roadmap

Semiconductor Hardware Roadmap

Process Node Evolution

Current and Future Nodes
| Node | Status | Key Players | Transistor Type |
|------|--------|-------------|-----------------|
| 5nm | Production | TSMC, Samsung | FinFET |
| 3nm | Production | TSMC, Samsung | FinFET/GAA |
| 2nm | Development | TSMC 2025, Intel 2024 | GAA |
| 1.4nm | R&D | TSMC 2027-2028 | GAA |
| Below 1nm | Research | Exploring CFET, 2D materials | TBD |

What "7nm", "5nm", "3nm" Mean Today
Node names no longer correspond to physical transistor dimensions. They primarily indicate:
- Density: Transistors per mm²
- Performance: Speed improvements
- Power: Efficiency gains

Transistor Architecture Evolution

Planar → FinFET → Gate-All-Around (GAA) → CFET (future)
        (16nm)     (3nm/2nm)              (sub-1nm)

AI Chip Capacity

NVIDIA GPU Production
| GPU | Process | Foundry | Supply Status |
|-----|---------|---------|---------------|
| H100 | TSMC 4N | TSMC | Supply-constrained |
| H200 | TSMC 4N | TSMC | Ramping |
| B100 | TSMC 4NP | TSMC | 2024 launch |

AI Accelerator Landscape
| Company | Chip | Status |
|---------|------|--------|
| NVIDIA | Blackwell | Upcoming |
| AMD | MI300X | Production |
| Intel | Gaudi 3 | Announced |
| Google | TPU v5 | Production |
| AWS | Trainium 2 | Coming 2024 |
| Cerebras | WSE-3 | Production |
| Groq | LPU | Production |

Capacity Constraints
- Leading-edge capacity: Limited to TSMC, Samsung, Intel
- Advanced packaging: CoWoS, HBM supply bottlenecks
- HBM memory: SK Hynix, Samsung, Micron; supply-constrained
- Geopolitical factors: US-China tensions affecting supply chains

Data Center GPU Demand
Estimated AI accelerator demand growing 30-40% annually, with supply lagging demand through 2025.