AI summary by TwinMind
Can You Do 7nm Chips Without EUV?
Asianometry
16m 32s
EUV vs DUV Lithography Technology Overview
- EUV (Extreme Ultraviolet) lithography machines cost $150 million each
- Companies can achieve 7nm without EUV but with significant compromises in performance and cost
- Prior to EUV, leading edge technology was DUV (Deep Ultraviolet) lithography
- Uses argon fluoride (ArF) laser with 193nm wavelength
- First introduced with ASML's PAS 5500 series around 2000
- Could etch 90nm node generation initially
- Improved to handle 65nm process node generation
Evolution of Lithography Technology
- Industry progression over 30 years: G-line mercury lamps → I-line → excimer lasers
- More powerful light sources improved resolutions and chip performance
- Planned smooth transition from 193nm to 13.5nm EUV light
- EUV missed delivery targets for 10 years, forcing industry to extend 193nm technology
Immersion Lithography Introduction
- Industry introduced immersion lithography to extend 193nm capabilities
- Shoots 193nm light through water to improve resolution
- ArFi (Argon Fluoride immersion) entered industry ~2007 with ASML XT 1700i machine
- Enabled 45nm process node
Node Naming Convention Changes
- Originally node names corresponded to smallest half pitch or gate length
- Moore's Law guides numbering: each generation 70% smaller than previous
- Progression: 65nm → 45nm → 32nm (2009) → 28nm (2010-2011)
- ArF lasers physically cannot achieve half pitch smaller than 40-45nm in single exposure
- Industry kept shrinking headline numbers by 70% for marketing despite unchanged half pitches
- TSMC terminology: N28, N10, N7, N5 (corporate communications standard)
FinFET Transistor Technology
- N22 process node (2012): industry transitioned to FinFET transistors
- FinFETs are 3D structures with multiple fins side-by-side covered by gate
- Three-gate FinFET up to 2x as effective as planar gate using same space
- Intel first to implement FinFET with N22 Ivy Bridge CPUs (2012)
- Samsung and TSMC followed later
- TSMC skipped from N22 to N16 when implementing FinFETs
- 3D structure interfered with multiple patterning techniques
Multiple Patterning Techniques
Double Patterning (Litho-Etch-Litho-Etch)
- Started at N22 node process generation
- Uses two lower density layout masks
- Two exposures superimposed to create single finer image
- Pitch split technique for dense layouts
- Calibre multi-patterning tool handles design splitting automatically
- Worked fine for N20, N16, N14 nodes
Triple Patterning Challenges
- Required for N10 node when EUV still unavailable
- Three masks of lower density layouts superimposed
- Introduces immense new complexities in EDA tools, foundry design rules, chip designs
- Computational problem: finding optimal way to split design into three masks
- Square with 4 nodes: 2 double patterning options vs 18 triple patterning variations
- 5 node structure: 30 variations
- NP-complete problem (not solvable in polynomial time)
- Chip designers required retraining
- Economic impact: double patterning increased lithography costs 50% from N28
- Triple patterning makes lithography module 3.85x higher cost than N28
N7 Node Implementation
- Final node achievable with 193i technology
- 70% node shrink from N10
- TSMC market leader by this point
- EUV close to ready but TSMC needed more time for comfort
- N7 hit high volume manufacturing 2018 (Apple A12, Huawei Kirin 980)
- Two versions: N7 (multiple patterning) and N7+ (EUV)
- Later transitioned to N6 node
Self-Aligned Quadruple Patterning (SAQP)
- N7 requires triple patterning + self-aligned quadruple patterning
- SAQP is double application of self-aligned double patterning (SADP)
- Process: template layer → lithography → CVD spacers → strip template → etch
- Density doubled, then process repeated (quadruple patterning)
- Creates high density parallel lines
- Extremely restrictive design constraints
- Fixed width parallel lines only
- Less efficient designs
- Extremely tight error margins
Intel's N10 Struggles
- Intel's N14 roughly equivalent to TSMC's N10 in transistor density
- N14 broke Intel's tick-tock strategy
- Launched 2014, remained flagship until N10 ramped 2019
- Intel's N10 roughly equivalent to TSMC's N7
- Attempted N10 without EUV using extreme multiple patterning
- Required quad, penta, hexa patterning for select features
- Up to 6 exposures needed to draw single feature
- Yield issues led to delayed market entry
EUV Breakthrough and Impact
- EUV became commercially viable 2014-2016
- ASML's dual shot laser pulse technique drastically improved power output
- N7 with EUV: 30% better cost efficiencies despite $150M machine cost
- 2020: TSMC celebrated 1 billion N7 chips (rare blog post)
Design Technology Co-Optimization
- Modern node processes worked alongside chip designs
- Everything optimized and nuanced: fin sizes, pitches
- Some features shrunk, others not
- Carefully curated design choices with trade-offs
- Node specifications fuzzy between generations and foundries
- "7nm" achievements vary significantly between companies
- Headlines don't reflect actual technical complexity
