AI summary by TwinMind
How Semiconductor DRAM Went 3D
Asianometry
19m 53s
Introduction and DRAM Basics
- Intel introduced first commercial Dynamic RAM (1970) - kicked off race for larger memory capacities
- Dynamic RAM memory cell stores 1 bit as charge in capacitor
- Capacitor = 2 conducting plates (electrodes) separated by insulating material (dielectric)
- Charge mapping: capacitor holds charge = 1, no charge = 0
- Single access transistor writes/reads charge
- Read operation destroys charge, requires refresh
- Refresh periods: few dozen milliseconds due to charge leakage
- 1-transistor design invented by Robert Denard at IBM
- vs SRAM: 6 transistors (2 cross-coupled inverters + 2 access transistors)
- Bare bones structure enables maximum density packing
Early Memory Scaling and Moore's Law
- Intel 1103 chip (mid-1970): 1 kilobit, revolutionized semiconductor memory
- First semiconductor memory with high storage, good performance, cheap price
- Intel set product cadence to Moore's Law
- Early memory Moore's Law: 4x bits per chip every 3 years
- Leading edge grants cost advantage per bit
- Memory as commodity: cost edge = significant economic advantage
Capacitor Physics and Scaling Challenges
- Memory cell scaling achieved by shrinking 1-transistor cell
- Read operation: charge flows from capacitor → access transistor → bit line
- Bit line voltage change indicates 1 or 0
- Charge must be detectable despite background radiation
- Capacitance critical for data integrity
- Capacitance equation: C = K × A/D
- K = dielectric constant
- A = electrode plate area
- D = distance between plates
- Early scaling: planar transistors/capacitors (capacitor lies flat)
- 64 kilobit (1982) → 1 megabit (1985): scaling via planar cell shrinkage
- Problem: shrinking capacitor = decreased plate area = reduced capacitance
- Solution: shrink distance between plates
- 1 megabit generation: tactic hit upper limit
Transition to 3D Structures
- Need for 3D transistor structure for 4 megabit generation
- Two design contenders: stacked vs trench capacitors
Stacked Capacitor Development
- Conceived 1976 by Mitsumasa Koyanagi (Tohoku University, working at Hitachi)
- Stack capacitor on top of access transistor and bit lines
- 1977: first test chip fabricated (3 micrometer process)
- 1979: 16 kilobit DRAM using stacked capacitor
Trench Capacitor Development
- 1974: Hideo Tsunami (Hitachi) attended TI conference on solar cells with deep trenches
- Tsunami's background: photo emission spectroscopy, radio enthusiast
- Recognized trimmer condenser concept (variable capacitor)
- Invented trench capacitor: capacitor below access transistor
- 1975: Japanese patent (no overseas patents)
- 1982: Hitachi presented trial 1 megabit DRAM using trench cell
4 Megabit Generation (1986-1990): Industry Split
Trench Capacitor Advantages
- 4 sides + bottom = increased plate size = better capacitance
- Hole structure = higher cell density
Trench Capacitor Disadvantages
- Manufacturing challenges: anisotropic etch (reactive ion etch) - immature technology
- Uniformity issues with insulator layer
- Stability problems
- Metrology difficulties (measuring progress in holes)
- Data integrity concerns:
- Current punch-through between trenches
- Soft error problem: alpha particles from cosmic rays change stored charge
- First trench variants particularly susceptible
Industry Adoption Split
Stacked Capacitor Adopters:
- Fujitsu, Mitsubishi, Hitachi (despite employee Tsunami inventing trench)
- Korean makers: Samsung, Gold Star (LG), Hyundai
- Hitachi chose stacked due to mainframe business concerns about soft errors
- Samsung chose stacked: easiest to manufacture, time-to-market paramount
Trench Capacitor Adopters:
- NEC, Toshiba, IBM, Siemens, Texas Instruments
- IBM created substrate plate trench (SPT) variant
- Filled trench with polysilicon (one plate)
- Silicon substrate = second plate
- Charge stored in trench itself
- Simplified wiring, overcame challenges
- Siemens, Toshiba adopted SPT through IBM research alliances
Market Outcome and Embedded Applications
- Stacked capacitor won due to superior manufacturability
- Most trench adopters dropped out of standalone DRAM business
- IBM continued trench technology for embedded DRAM (eDRAM)
- eDRAM: memory cells on system-on-chip (SoC)
- Offers more capacity/data transfers than on-chip SRAM cache
- Trench density advantages relevant in space-constrained SoCs
- Manufacturing benefit: fully planarized (same horizontal plane as logic transistors)
- Helps lithography with shallower depths of focus
16 Megabit Generation (1988)
IBM SPT Approach
- Added thick (100 nanometers) insulating collar of silicon dioxide around trench
- Prevented neighboring trench interference
Stacked Capacitor Evolution
- 1988 Fujitsu proposal: fin capacitor modification
- Cross-section resembles palm tree
- Fins made from polysilicon (serve as plates)
- Dielectric between fins
- Complex manufacturing: multiple sacrificial layers (temporary layers laid down then etched away)
Material Innovations: Hemispherical Grain Structure (HSG)
- Critical innovation for 64 megabit and 256 megabit generations
- Polysilicon grains added to stack capacitor's bottom plate
- Created textured/rugged surface = increased plate surface area = higher capacitance
- HSG advantage for stacked capacitors: trench capacitors couldn't implement (heat baking compromised structural stability)
- NEC invented HSG (late 1980s/1990)
- Micron Technology first ramped up (1997)
- NEC, Samsung introduced into 64 megabit products (1998)
- 2001: 70% of DRAM makers adopted HSG
Gigabit Era Innovations
Cylindrical Capacitors
- Late 1990s: 193 nanometer DUV lithography enabled higher resolutions
- Cylindrical capacitor concept: 1989 VLSI Symposium (Mitsubishi team)
- Evolution of fin capacitor, maximizes plate surface areas
- Samsung 90nm node (2005) 512 megabit DRAM details:
- Polysilicon cylinder etched using deep reactive ion etching
- Two-layer dielectric: aluminum oxide + hafnium oxide (inside/outside cylinder)
- Second plate: crystallized silicon inside cylinder
- Titanium nitride layer (atomic layer deposition)
- HSG application for additional surface area
Exotic High-K Dielectric Materials
- Industry progression:
- Started: silicon dioxide
- 1 megabit: abandoned silicon dioxide
- 4 megabit → 256 megabit: stacked dielectric (silicon nitride + silicon dioxide)
- Gigabit era: entirely new materials
- Tantalum pentoxide (first exotic substance):
- Industry already used tantalum in manufacturing
- Dielectric constant: 25 (vs 6-7.5 for previous composite)
- 3x improvement = useful for only one generation
- Super exotic materials: barium strontium titanate
- Required new deposition methods (metal organic CBD)
- Heavy reliance on exotic materials due to limited UV lithography availability
Industry Reflection
- DRAM's 3D evolution more aggressive than logic side
- Logic: moved to FinFET (2011), gate-all-around (15 years later)
- DRAM's commodity nature forced aggressive pursuit of 3D capacitor structures
- Tough business driving ridiculous complexity in advanced memory semiconductors
- Memory often underappreciated ("chop liver") of semiconductor industry
