AI summary by TwinMind
The 3-D Transistor Transition
Asianometry
14m 18s
Historical Context and Industry Transition
- 2011: Intel and leading semiconductor industry made remarkable technical transition to 3D transistors
- First time producing transistors in three dimensions - thin field effect transistor (FinFET)
- Transition took over 10 years in the making
- Industry building block for decades: metal oxide semiconductor field effect transistor (MOSFET)
- Common in digital circuits: inverters, NAND gates, SRAM cells
- Components: gate sitting on top of channel connecting source and drain
- Source/drain: silicon regions doped with other elements to donate/receive electrons
- Thin insulating layer between gate and channel (usually silicon oxide)
- Gate doesn't cut through channel - whole structure is flat/planar
Classical Scaling Era and Dennard Scaling
- First few decades: new process nodes delivered performance, power, area gains by shrinking transistor dimensions
- Classical scaling: cramming more transistors onto chip
- Benefits: electric signals had less distance to travel between transistors
- 1974: IBM researcher Robert Dennard identified secondary benefit - smaller transistors used less power
- Moore's Law: reduce transistor linear dimensions by half every 3 years
- 1980s: dimensions fell under 1 micron for first time
Short Channel Effect and Industry Crisis
- Mid-1990s (350nm process): transistors started acting "funny"
- Physical dimensions shrinking caused problems:
- Source and drain getting closer together
- Insulating layer getting thinner (1.2nm at one point - 5 atoms wide)
- Channel itself getting thinner
- Gate's control over current flow weakened
- Short channel effect: current "dives under the gate" like rabbit burrowing
- Current could sneak from source to drain even with gate closed
- Could travel through channel part furthest from gate or through silicon substrate
- Serious power consumption issues - leakage meant smaller transistors didn't follow Dennard scaling
- Transistors would consume as much energy in off state as during on state
- Consumer electronics becoming more portable - higher power efficiency demands
Industry Response and DARPA Initiative
- Classical MOSFET structure had final endpoint - practical size limit at ~20nm
- 1996: Leading edge at 259nm, DARPA aware industry had no long-term plan beyond 2002
- DARPA called for proposals on sub-25nm devices - "25 nanometer switch"
- Received 10-12 proposals from IBM, AT&T, Stanford
- All were mere extensions of original MOSFET structure
- DARPA wanted something ambitious and feasible
UC Berkeley's Winning Proposal
- Winner: University of California Berkeley team led by Professor Hu Chenming
- Hu joined Berkeley 1976, initially studied energy/hybrid cars
- Government funding dried up after Reagan administration, pivoted to semiconductor research
- 1982: sabbatical at National Semiconductor, saw coming end of planar transistor
- Proposal created over span of one week
- Title: "Novel fabrication device structures and the physics of 25 nanometer FETs for terabit scale electronics"
Two Key Ideas from Berkeley Proposal
Fully Depleted Silicon on Insulator (FDSOI)
- Problem: electrons inappropriately traveling using silicon substrate
- Solution: add new insulation layer on top of silicon substrate
- Became moderately successful for specific uses (analog, power electronics)
FinFET Concept
- Sketched by Professor Hu during long flight
- Focus: allow gate better control over channel
- Traditional planar: source, drain, channel are flat with gate on top
- FinFET: turns planar structure onto its side, rises above surrounding area
- 3D shark fin appearance gives it name
- Gate wraps around top and sides of channel vs just sitting on top
- Primary benefit: gate wraps around channel on three sides vs one side for planar
- Smaller physical footprint - can stuff more onto same real estate
Historical Precedents
- Idea not entirely new - two similar proposals made before but never implemented
- 1990: D. Hisamoto at Hitachi proposed wrapping gate around channel
- Called it "fully depleted lean channeled transistor" or Delta
- 1980s: Texas Instruments paper called for "trench transistor"
- Concept useless without execution - nobody knew if could scale to 25nm or smaller
DARPA Development Phase
- 4 years prototyping working FinFET device with DARPA financial backing
- Team had freedom for long-term goal + ability to draw on private/public resources
- Used equipment/facilities at Stanford and Lawrence Berkeley National Lab
- 2001: DARPA project completed, team published results
- FinFET recognized as tremendously innovative but not too disruptive
- Traditional lithographic and etching tools could still be used
- Industry could get better performance/power efficiency without abandoning billions in equipment/decades of experience
Industry Adoption Delays and Interim Solutions
- Conservative semiconductor industry didn't immediately adopt FinFET
- Opted for short-term measures to extend MOSFET life
Strain/Stress-Based Engineering (90-45nm nodes, early 2000s)
- Added layers of "strained silicon" to channel
- Usually on top of silicon germanium buffer layer
- Strained = silicon crystal atoms being stretched
- Helped electrons travel faster across channel
- Required growing strained silicon layers using epitaxy
High-K Metal Gate (28nm node, 2009-2010)
- Replaced silicon dioxide insulating layer with high-K metal (hafnium oxide)
- High-K (relative to silicon dioxide) kept opposing charges apart
- Led to "gate first vs gate last" dilemma splitting industry
- Caused semiconductor manufacturers to drop off leading edge
- Benefit was one-time only - needed new solution after 28nm
Intel's FinFET Implementation
- 22nm node: Intel implemented 3D device called "tri-gate transistor"
- Descendant of original FinFET
- Original had two gates on sidewalls (double gate FinFET)
- Intel's version: three gate layers on all three exposed sides of fin
- Most people still refer to them as FinFETs
Manufacturing Challenges
- Prototype to mass production always challenging
- Critical parameters:
- Fin geometry (height, etc.)
- Fin pitch (spacing between fins) ~60nm
- 22nm process: fin just 10nm wide
- Millions of nanoscale fins fabricated with extremely little variation
- New multi-patterning techniques required
- All foundries struggled with scaling first FinFET products, yield issues, delays
Industry Adoption Timeline
- 2013: TSMC and Samsung jumped to FinFET (16nm and 14nm processes)
- 2014: GlobalFoundries joined with 14nm process (licensed from Samsung)
- FinFET worked well through N7 and N5 generations
- Recently losing effectiveness at smaller nodes
Current Transition to Gate-All-Around
- Industry decided on gate-all-around as new architecture
- FinFET worked by covering more surface area - gate-all-around takes further
- Completely surrounds channel including bottom
- Can stack multiple channels/nanowires on top of each other through gate
- June 2022: Samsung started shipping 3nm process with gate-all-around
- TSMC sticking with FinFETs for N3 but adopting gate-all-around for N2 (under construction in Baoshan)
Economic Challenges
- 3D transistors are technical marvels but expensive
- FinFET's intricate construction: 28nm (last with planar gates) = point where cost per gate stops declining, starts rising
- Gate-all-around accelerates cost trend further
- Challenge: building gaps between nanowires when can only etch downward
- Solution involves super lattices and atomic layer deposition methodology
- Leading edge becoming economically infeasible for all but biggest companies
- ASML CTO: reaching end of lithography not due to technology but economic feasibility
- Can always develop next structure but question is who can pay for it
