AI summary by TwinMind
An Infamous Transistor Dilemma: Gate First or Gate Last?
Asianometry
22m 11s
MOSFET Fundamentals and Early History
- MOSFET = metal oxide field effect transistor with 4 major parts: gate, source, drain, channel
- Source/drain = silicon regions doped with elements to donate/receive electrons
- Gate activation creates electric field attracting electrons, forming conducting layer between source/drain
- Gate oxide = thin pure oxide film (50-5nm thick) on top of channel, most challenging part to manufacture
- Late 1960s: aluminum metal gates + silicon dioxide gate oxide (explaining "MOS" name)
- Aluminum melting point 660°C problem: had to diffuse source/drain at 1000°C+ before gate production
- Gate done last, alignment errors caused shorts and excess power consumption
Polysilicon Gate Revolution
- Fairchild, Intel, others replaced aluminum with heavily doped polysilicon
- Enabled gate-first process: gate used as reference to align source/drain
- Self-aligned polysilicon gate technology became mainstream for 40+ years
- First MOSFET: 20 micrometers, gate oxide 100-150nm thick
Scaling Challenges and Short Channel Effects
- 130nm node (2001): 70nm channel length, 20-25nm gate oxide thickness
- Short channel effects emerged: source/drain electric fields weakened gate control
- Electrons jumped over gate causing leakage and power consumption
- 90nm node (2004): 45-50nm channel, 1nm gate oxide thickness
Intel's Strained Silicon Innovation (2002)
- Intel shocked industry by adding strained silicon to 90nm node (competitors planned for 65nm)
- Strained silicon = crystal lattice stretched beyond normal distance
- Electrons/holes travel faster through strained vs traditional silicon
- CMOS implementation:
- P-MOSFET: silicon germanium trenches near gate (compression strain)
- N-MOSFET: silicon nitride layer on top (tension strain)
- Claimed 2% wafer cost increase but combined with low-k dielectric caused delays
- Pentium 4 Prescott delayed months, thermal issues, modest performance gains
45nm Node and High-K Transition
- Mid-2000s smartphone revolution demanded power efficiency
- Gate capacitance formula: C = K × ε × area / oxide thickness
- Area shrinking, ε constant, oxide thickness at limit (1.2nm = 11 hydrogen atoms)
- Only variable: dielectric constant K
- Silicon dioxide K limitations led to research of high-K materials
High-K Material Research and Development
- Industry researched titanium dioxide, hafnium silicate, zirconium silicate
- Hafnium oxide chosen: K=20-25, high temperature stable, silicon compatible
- Atomic layer deposition enabled precise hafnium oxide deposition
- Problem: hafnium oxide reacted with polysilicon gate at high temperatures
Gate First vs Gate Last Approaches
Gate First (IBM Alliance Method)
- Metal Inserted Polysilicon Structure (MIPS)
- Titanium nitride layer over hafnium oxide + polysilicon gate + silicide
- Build gate before high temperature source/drain processes
- Championed by Sematech, IBM Research Alliance (IBM, Samsung, Chartered)
- Simpler integration, similar to existing flow
- Risk: high temperature damage to gate integrity
Gate Last (Intel Method)
- Deposit hafnium oxide + polysilicon gate for alignment
- Destroy/remove gate materials after source/drain creation
- Rebuild gate with aluminum/metals in resulting hole
- Different recipes for NMOS/PMOS transistors
- 4% wafer cost increase, 10% density penalty
- Chemical mechanical polishing (CMP) added complexity
- 2009: Intel developed "High-K Last" second generation
Intel's 45nm Success (2007)
- Oregon and Chandler fabs began high volume production H2 2007
- First in industry with high-K, became highest yielding process
- Delayed 193nm immersion lithography adoption until 32nm node
- Gordon Moore called it "biggest change since polysilicon gate MOSFET introduction"
Foundry Competition and Market Dynamics
- TSMC focused on 193nm immersion for 45nm, used triple gate oxide
- 193i delays led to 40nm relabeling, competitive pressure from SMIC
- IBM Alliance announced 32nm high-K gate first (April 2008)
- TSMC responded at 2008 Technology Symposium, committed to gate first for 32nm high-performance
TSMC's 40nm Struggles and Pivot
- 40nm node delayed ~1 year (Q1 2009), affected Nvidia GeForce 100 series
- Rick Tsai lost CEO position due to delays
- 40nm eventually successful: 32% of 2012 revenues, 5-6% of 2023 revenues
- 32nm node killed due to poor yields
- Focus shifted to 28nm node
Industry Transition to Gate Last
- TSMC ran parallel gate first/last development for 28nm
- February 2010: chose gate last for better design flexibility
- Different metals for NMOS/PMOS enabled threshold voltage adjustment
- Morris Chang predicted gate first companies would switch at 22nm node
- Samsung initially used gate first for 32nm/28nm, secretly developed gate last R&D
Final Industry Convergence
- 2011: IBM and Global Foundries abandoned gate first, switched to gate last for 20nm
- Gate first stability/reliability issues with thermal processes
- Samsung switched to gate last as well
- TSMC's 28nm success: high-K metal gate + 4th gen strained silicon + 193nm immersion + copper interconnect
- Considered pinnacle of planar transistor technology
Intel's FinFET Transition
- 2011: Intel announced FinFET transition, making high-K metal gate "old news"
- High-K remains core part of Intel innovation legacy
- Demonstrated Intel's technological leadership: insight, execution, global scaling
- 3 years after high-K breakthrough, moved to FinFET
- Remarkable technological lead eventually squandered
