AI summary by TwinMind
Nvidia's Computational Lithography Breakthrough
Asianometry
15m 10s
Nvidia CuLitho Introduction
- March 2023: Nvidia announced CuLitho computational lithography library
- Claims 40x performance leap beyond current lithography
- ASML and TSMC endorsing the technology
- Presentation by Dr. Vivek K Singh: "Accelerating Computational Lithography Enabling Our Electronic Future"
CuLitho Technical Overview
- Collection of parallel algorithms by Nvidia scientists
- Accelerated primitive operations for OPC (Optical Proximity Correction)
- Algorithms accelerated hundreds of times using GPUs
- Integrated into traditional OPC EDA software using parallelism
- GPUs process algorithms faster than CPUs
- Overall system speedup: 23-42x end-to-end computational lithography
Lithography Background & Challenges
- Optical lithography: multi-million dollar camera projecting light through/off masks
- Until late 1990s: feature sizes larger than light wavelengths
- 180nm node generation: used 248nm KrF DUV laser
- 193nm ArF laser available but not widely used
- Problem: small feature sizes cause pattern transfer errors when resolution/image quality deteriorates
Rayleigh's Equation & Resolution Enhancement
- Resolution = K1 × wavelength / numerical aperture (NA)
- Options: raise NA or lower K1
- Raising NA trade-offs:
- Collects more light (immersion lithography concept)
- Thinner depth of focus (vertical distance for acceptable image quality)
- Requires thinner resist layer
- Compromises etch process protection
- Lowering K1: represents manufacturing process contribution (resist, materials, imaging techniques)
Computational Lithography History
- Early 1970s: began quantifying lithography principles
- Previously artisan technique dependent on operator skill
- IBM's Dr. Rick Dill led efforts creating computer models
- Models simulate resist-coated wafer surface reaction to light
- Early 1990s: computational lithography became useful for real-world applications
Resolution Enhancement Techniques (RETs)
- Optical tricks to improve design transfer to wafer
- Three major RETs introduced 25+ years ago for 180nm node:
- Phase shifting masks
- Off-axis illumination
- Optical Proximity Correction (OPC)
Optical Proximity Correction Evolution
- Modifications to chip design layout compensating for optical distortions
- "Misprint" design onto quartz photomask so light prints correctly on resist-coated wafer
- Mid-1990s: simple geometric rules-based OPC
- Adding tick marks (serif-like) at design line ends
- Used for metal interconnect layers higher in IC
- Model-based OPC emerged as feature sizes shrank
- Extensive computational lithography simulations
- Corrects entire layout before fab
- Process: GDSII file → OPC software → add corrections
- Breaks polygons into segments, samples within segments
- Simulates light journey through sampled mask segments
- Iteratively optimizes mask design
- Very accurate but computationally intensive
Inverse Lithography Technology (ILT)
- 1981: Professor B.E.A. Saleh (University of Wisconsin-Madison) proposed concepts
- Different approach from traditional OPC:
- Traditional OPC: mask → wafer (forward)
- ILT: wafer → mask (inverse/backward)
- Starts with desired wafer pattern, works backward to optimal mask design
- Writes entire mask pixel-by-pixel from scratch
- Professor Saleh used simulated annealing optimization
- Process: random mask generation → random pixel changes → keep improvements
ILT Development Timeline
- Post-Saleh: Berkeley, IBM, other organizations refined techniques
- 2003: Luminescent Technologies startup
- Paper: "Fast Inverse Lithography Technology"
- Gave ILT its present name
- 2007: Intel presented their version
- Called it "pixelated mask technology" (avoided ILT name)
- Intel funded academic research since 1990s
ILT vs Model-Based OPC Comparison
- ILT advantages:
- More accurate results
- Greater fidelity to desired pattern
- More room for process error tolerance
- ILT disadvantages:
- Massive computational power requirements
- Some designs take weeks to process
- Manufacturing challenges with curvilinear masks
- Complex curves/diagonals vs traditional Manhattan geometry
- Electron beam mask writing slow but precise
- Fracturing complex shapes into rectangles challenging
ILT Market History
- Late 2000s/early 2010s: ILT too early for widespread adoption
- Luminescent pushed ILT for 45nm and 32nm nodes
- 193nm immersion lithography improved resolutions enough that manufacturers didn't need ILT over OPC
- 2012: Luminescent split up
- Parts acquired by KLA (equipment maker)
- Parts acquired by Synopsis (EDA maker)
- EDA vendors adopted ILT for special situations, not widely adopted
Modern ILT Revival
- Mask writing tools improved over time
- Multi-beam tools from Austria-based IMS enable reasonable curvilinear mask production times
- Computational challenges remain significant
- Modern chip complexity: Hopper has 80 billion transistors (nearly double previous generation)
- Trillions of polygons on single chip
- Increasing demands forcing fabs to use ILT more in advanced nodes
- TSMC presented simulation results showing GPU potential for ILT speedup
- Nvidia accepted challenge, resulting in CuLitho
Additional Computational Lithography Techniques
- Source mask optimization: optimize both mask pattern and illuminating light source
- Massive field with many functionalities beyond video scope
- Self-referential nature: using computers to make chips for faster computers
