AI summary by TwinMind
Why Every 3nm Process Node is Different
Asianometry
14m 4s
TSMC N3 Process Node Variants
- 2022 Technology Symposium announcement of N3 process node (marketed as 3nm)
- Multiple N3 variants planned: N3, N3E, N3P, N3X
- Reflects new reality: no single node does everything
- Approach: pick and co-develop right node for specific product requirements
Moore's Law Glory Days (1980-2005)
- 70% shrinkage of chip features every 2 years
- Simple scaling in both X and Y directions
- Designers created chips without considering fabrication constraints
- Foundries handled shrinkage automatically
Lithography Limitations Era (2005+)
- 2005: foundries hit fundamental limitations
- 2001: 193nm laser came into production - essentially end of line
- 157nm lithography failed to take off
- EUV still years away
- 130nm node: first to encounter problems
- High defect rates from inadequate lithography
- Previously: random/unavoidable defects (particles) were main issue
- New issue: lithography-related errors gaining ground
Resolution Enhancement Techniques
- Response to lithography struggles
- Added tick marks at end of design lines (like serif fonts)
- Made design lines bigger overall, easier to print
- Introduced "guard rails" at chip design stage
- Goals: avoid killer defects, achieve best yields
Power Performance Area (PPA) Metrics
- Industry yardstick for chip evaluation
- Power: consumption during usage/idling
- Performance: maximum attainable frequency
- Area: silicon area in square millimeters (or gate count)
- Recent addition: Cost (PPAC metric)
- Mobile market prioritizes low power usage and smaller areas
Design for Manufacturability (DFM)
- First design stage guard rails
- Arrived with 90nm node
- Derived from resolution enhancement techniques from 130nm generation
- Primary goal: work around 193nm lithography shortcomings
- Key issue: corners never truly corners at nanoscale
- Lithography physically incapable of making true corners
- Always some curvature interfering with chip design
- New rule: cannot put corner too close to transistor gate
- Trade-offs: conservative approach = better yields but unused space/performance hit
- 90nm node: 4x more design rules than 180nm node (just 2 generations prior)
- 65nm and beyond: rulebooks kept growing
- 193i immersion lithography didn't slow rule growth
- Problems: rules introduced with little warning, one-way road of restrictions
Design Technology Co-Optimization (DTCO)
- Emerged around 28nm process node
- Tighter relationship between foundries and chip designers
- Began with N20, N16, N14 nodes
- Two key innovations: FinFET transistors and multi-patterning
FinFET Technology
- 3D transistors replacing planar MOSFETs
- Planar: gate sits between source and drain (flat structure)
- FinFET: raises source/drain structure higher (3D fin)
- Fin dimensions (height, width, length) affect power/performance/area/cost
- Thinner/higher fins perform better, fit more in space
- Risk: fin collapse or flop over
DTCO Process
- Chip designer works directly with foundry process engineers
- Goal: process node meeting project's PPAC targets
- Step 1: specify exact parameters (fin thickness, height, distances between fins)
- Still guided by 70% scaling targets from Moore's Law
- Trade-offs across all four PPAC metrics
- Example: low-power mobile chip prioritizes power and area metrics
Standard Cell Library Creation
- Process engineers craft standard cell libraries from fin specifications
- Standard cells: standardized groups of gates enabling logical functions (AND, OR, etc.)
- Lego blocks of chip design, made up of fins
- Sized with track heights metric
- Higher track heights: more power, more area
- Lower track heights: opposite trade-off
Scaling Example (N14 to N10)
- Goal: 0.47x total area scaling for particular standard cell
- N10 cell covers 0.47 of N14 cell area
- Achievement breakdown:
- Design contribution: 9 tracks to 7.5 tracks = 0.83x area scaling
- Manufacturing contribution: triple patterning = 0.56x area scaling
- Combined: 0.56 × 0.83 = 0.47x final goal
DTCO Implementation Characteristics
- Process starts years before high volume manufacturing
- Creates very different leading edge process nodes between foundries
- Requires significant trust between foundries and customers
- Information sharing: chip designers reveal design intentions, foundries share process complexity
- Unlike DFM: foundry accepts more process complexity for designer flexibility
- Bilateral trust and information flow
Future: System Technology Co-Optimization (STCO)
- IMEC exploring next evolution beyond DTCO
- Extends AMD's chiplet approach
- Brings advanced packaging industries into collaboration circle
- Goal: segregate parts that cannot shrink (I/O, analog components)
- Uncertain if DTCO will mature into STCO
- Need to account for previously overlooked functions for more PPAC gains
Industry Integration Trends
- DTCO resembles old vertically integrated semiconductor makers (IDMs)
- Past 30 years: academic/business discussion of industry split into two sides
- Reality: leading edge foundries and designers working closer than ever
- Contradiction between theory and practice
