AI summary by TwinMind

TSMC's First Breakthrough: The Copper/Low-K Interconnect Transition

Asianometry

34m 50s

Introduction to Interconnects and RC Delay Problem

  • Interconnects = wires transmitting electrical signals between transistors and circuit elements
  • Traditional materials: aluminum interconnects + silicon dioxide insulating layers (30+ years)
  • Late 1990s: technical necessity for new materials due to RC delay issues
  • RC delay = resistance + capacitance causing signal propagation delays
    • Resistance: impedes electron flow, proportional to wire length
    • Capacitance: unwanted charge storage between closely spaced wires
    • Combined effect: extra time for signals to travel through chip

Interconnect Types and Manufacturing Process

  • Metallization = back-end process after transistor fabrication (front-end)
  • Three interconnect types:
    • Local interconnects: connect adjacent elements, thinner/closer spacing, heat resistant materials (polysilicon/tungsten/aluminum)
    • Global interconnects: span large chip portions, thick/low resistance for power/clock signals
    • Medium-distance interconnects: semi-global/inter-module connections
  • Memory chips: 3-5 metal layers (repetitive, dense structure)
  • Logic chips: 10-15 metal layers (irregular, requires more connections)

Traditional Aluminum-Silicon Dioxide Process

  • Aluminum: 4th lowest resistance (after silver, copper, gold)
  • Silicon dioxide: easy chemical vapor deposition
  • Manufacturing process: deposit aluminum → lithography → etch unwanted parts
  • Aluminum + silicon dioxide reaction creates protective alumina layer
  • Process worked well for first 30 years until scaling limitations

Scaling Challenges Leading to Crisis

  • 1978: IBM M2 interconnect pitch = 20 micrometers
  • 1994: Global layer pitch = 1.8 micrometers (supposed to be widest)
  • Scaling methods created problems:
    • Shrinking interconnect size → higher resistance → more RC delay
    • Making wires taller → manufacturing difficulties at nanometer scale
    • Adding more metal layers → more process steps → worse yields
  • 250nm node (1996): last generation using aluminum
  • 100nm projection: 6x longer signal transmission vs transistor switching

Copper Solution Development

  • Copper interconnects: 35% less resistance than aluminum
  • Low-k dielectric: replace silicon dioxide (k=3.9) with lower capacitance material
  • Combined solution: potentially 4x RC delay reduction
  • September 1997: IBM announces successful copper interconnects after 15 years R&D
  • Project "Red Aluminum": team worked 18 hours/day in secret
  • First mass production: early 1998 for Apple PowerMac

Copper Manufacturing Challenges

  • Copper poisoning: atoms diffuse into silicon/silicon dioxide
  • Solution: tantalum-based diffusion barriers
  • Cannot etch copper (no suitable etchant)
  • Damascene method (named after Damascus metalwork):
    • Lay intermetal dielectric → lithography → etch trenches/vias
    • Apply diffusion barrier → fill with copper → polish excess
    • Reverse of traditional subtract process (additive instead)

Key Technological Breakthroughs

  • Late 1980s: tantalum/tantalum nitride diffusion barriers discovered
  • 1985: IBM invents damascene method
  • 1989: superfilling/bottom-up electroplating breakthrough
  • 1993: first multi-level copper wires demonstration
  • 1995: dual damascene methodology (trenches + vias together)
  • 1997: high-volume production commitment

Industry Response and Competition

  • Pre-1989: only IBM investigating copper
  • Post-1989: industry-wide recognition of aluminum limitations
  • Competitor timelines:
    • Motorola: 1990 (guessed through IBM PowerPC design changes)
    • AMD: 1995
    • AT&T Bell Labs: ~1995
    • Intel: 1997 (despite own 1989 electroplating research)
  • Sematech R&D consortium accelerated knowledge sharing
  • Two weeks after IBM announcement: Motorola announces copper success

180nm Node Implementation (1999)

  • Major technical changes: copper + 30% feature reduction + 193nm lithography
  • 300mm wafer transition delayed to 130nm node
  • 1998: IBM PowerPC 750CX (300-400MHz) first copper chips
  • Motorola: 220nm half-node (late 1998) for PowerPC G4
  • Intel: delayed copper until 130nm (November 2000) due to scaling needs
  • Asian foundries: started from behind, had to rapidly learn copper technology

TSMC vs UMC Foundry Battle

  • 1999 market share: UMC 30%, TSMC 27%
  • Capacity race: UMC 1.7M vs TSMC 1.9M 8-inch wafers
  • UMC CEO: "trends indicate we will pass TSMC in revenue next year"
  • Different copper learning strategies:
    • UMC: joined IBM research consortium
    • TSMC: independent development by 6-person elite team
  • 1999: both shipped 180nm with top 2 copper layers (vs IBM's 6/7 layers)

Low-k Dielectric Challenges

  • Three attempts at low-k materials:
    • 250nm: Fluorinated Silicon Glass (FSG) k=3.5 vs 3.9
    • 180nm: Hydrogen Sesquioxide (HSQ) k=2.8 via spin-on technique
    • 130nm: Multiple competing approaches
  • HSQ failure: passed internal tests but failed high-volume production
  • TSMC crisis: discovered failure at Christmas, worked through holidays
  • Texas Instruments suffered same HSQ problems

130nm Node Material Wars

  • Industry split over low-k dielectric choice:
    • IBM/UMC/Infineon: Dow Chemical "Silk" (k=2.6-2.2) via spin-on
    • Texas Instruments/TSMC: Applied Materials "Black Diamond" (k=2.8) via CVD
    • Intel: continued with FSG for capacity reasons
  • Dr. Zhang (TSMC): personally swore off spin-on technique after HSQ failure

TSMC's 130nm Victory

  • October 2001: first to ship 130nm with copper + low-k dielectric
  • Primary customers: Nvidia and others
  • Backup FSG dielectric option for risk mitigation
  • Fab 6 and 12 in Hsinchu for production
  • Yield progression: teens → 70%+ by late 2002
  • End 2001: UMC defected from IBM/Infineon, switched to Novellus CVD

IBM's Silk Failure

  • IBM failed to ship any product with Silk
  • 2002: Xilinx publicly blamed IBM Silk for FPGA delays
  • Xilinx president criticized IBM manufacturing skills
  • 2003: IBM CTO defended decision but eventually switched to CVD
  • Bernie Mayerson: "industry should be saying thanks guys, you saved us from this"

Strategic Impact and Recognition

  • TSMC's 130nm success established technological leadership
  • 2003: Taiwan government awarded R&D team outstanding scientific achievement
  • Industry perception shift: TSMC from "simple factory" to innovation leader
  • Set foundation for TSMC's future dominance in advanced nodes

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome