How Semiconductor Yields Vastly Improved

Asianometry

Historical Semiconductor Yield Problems

Decades of semiconductor fabs tolerating bad yields: <50%, sometimes 20-10%
1978 US-Japan semiconductor dispute data:
- Both American and Japanese 16k DRAM makers started at ~2% yields
- By 1984: American 16k DRAM yield improved to 36%, Japanese reached 48%
- 12% yield gap accounted for 80% of cost difference between Japan and US
Modern comparison: TSMC's N5/N4 process nodes (4-5 years old) achieve ~80% yields
- 6 micron process vs 4 nanometer: dramatic complexity increase yet higher yields

Manual wafer inspection through microscopes
Human effectiveness varied by factor of 5 even with same training
Inspector performance pattern:
- Started low each shift
- Peaked 1/3 into shift
- Declined thereafter
- Never exceeded 87% effectiveness
Particle defects <1 micrometer, barely visible with microscope
Subjective classification made data comparison impossible across time periods

First known automated mask inspection tool
Compared two die images side-by-side using two 3-micrometer laser beams
Limitations:
- Frequent misalignments from mechanical issues
- Temperature fluctuations caused false positives
- Cannot detect repeated defects
- Cannot determine fatal vs non-fatal defects
- Cannot detect defects <2 micrometers
Breakthrough: long-term consistency enabled statistical comparison across months/years
Used by AT&T's Western Electric but not commercially available

Founded by Bob Anderson and Ken Levy with $600k VC funding
Levy background: aerospace engineer → flight simulation → Computer Vision
Computer Vision diversification:
- Primary: computer-aided design (CAD/EDA software)
- Side business: semiconductor manufacturing equipment (40% of revenue at peak)
- Bought lithography alignment invention, acquired Coblet mechanics company
- Employed 350 people making alignment machines, photo resist coders
1975: Levy and CFO left to start KLA
Computer Vision sold equipment lines: wafer probe/coder to Tokyo Electron, rest to Applied Materials (1978)

First product: automatic photomask inspection system
Laser crawled across mask, recorded reflections
Performance improvement: 8 hours → 5 minutes for 3-inch mask inspection
KLA-200 series (early 1980s): added database reference image comparison capability

Lithography process example:
- Exposure tool: UV light exposes resist-coated wafer
- Development: hardens resist for etch step
- Etch: permanent cutting into substrate using hardened resist guide
- Inspection: check for bridges/gaps in design patterns, alignment verification

1990s: faster computers for image processing, stored chip design file comparison
2000s major transition: CCDs → CMOS image sensors
- Enabled much faster image capture
- Opened door to digital processing
Light source evolution: broadband plasma (BBP)
- Laser ignites plasma for intense broadband light
- Wavelength range: visible light to UV
Focus: contrast/signal-to-noise ratio rather than resolution
Recent: deep UV wavelengths for detecting defects as small as 10nm

Processes: post-wafer fabrication (assembly, test, packaging)
More lenient cleanliness requirements, lower prices
Rudolph Technologies: entered from microscope background
- Focus: bumps, redistribution layers, TSVs for advanced packaging
- 2019 merger with Nanometrics → Onto Innovation (publicly traded)
Camtech: familiar player in low-end automated optical inspection

Intel, TSMC, Samsung: deeply conservative companies
Process nodes run for years without changes
Reluctant to change even potentially beneficial technologies
Yield management: long, slow process requiring accumulated efforts of thousands of workers
Technical miracle combining fast microscopes with "blood, sweat, and tears"

Try for free