AI summary by TwinMind

Can SRAM Keep Shrinking?

Asianometry

14m 50s

TSMC N3 Node SRAM Issues

  • Early 2023: TSMC presented N3 (3nm) node details
  • Two N3 variants: N3B (dead end, mostly Apple) and N3E (Nvidia/AMD customers)
  • Critical issue: N3E SRAM bit cell same size as N5 nodes
    • No SRAM scaling from N4 to N3 generation
    • Significant industry concern over this revelation

Memory Hierarchy Fundamentals

  • Memory technologies sorted by response time
  • Fast end: Level 1/2/3 caches (SRAM on-chip)
    • Access data within few nanoseconds
    • Expensive to manufacture, takes precious silicon space
  • Slow end: Dynamic RAM, flash memory, hard drives
    • Cheaper, higher capacity, off-chip (motherboard)
    • More time/energy to access data
  • SRAM hugely important for IC cost and functionality

SRAM Historical Development

  • 1963: First SRAM design patented by Robert Norman (Fairchild)
    • Designed on friend's breakfast table in few hours
    • IBM requested patent, Fairchild held firm
    • Initial design used bipolar transistors
  • 1964: Fairchild patented MOSFET design (more scalable)
  • 1970: 256-bit TTL SRAM used in Burroughs ILLIAC 4 main memory
  • 1971: IBM System 370 Model 145 shipped with 128-bit bipolar SRAM
  • Since then: One of most widely manufactured memories

SRAM Technical Architecture

  • Volatile memory (loses data when power lost)
  • Unlike DRAM: No capacitor, no periodic refresh needed
  • Most common: 6-transistor SRAM cell design
  • Uses bistable latch with two stable states (0s and 1s)
  • Latch made of two inverters in feedback loop
    • Each inverter: 2 transistors (4 total)
    • Plus 2 pass gate transistors for read/write
  • Memory organization: Cells in rows connected to word lines and bit lines

SRAM Area Consumption in Modern Chips

  • High-performance SoCs demand better embedded memories
  • Mid-2000s: Some CPUs had 71% die area as SRAM
  • Today: Some systems 90% surface area covered in embedded memory
  • Portable devices: Less embedded memory (space/power constraints)
  • Systems always craving memory, but memory doesn't scale as fast

Power Leakage Challenges

  • Traditional focus: Dynamic power during switching
  • Modern focus: Leakage power as transistors shrunk
  • Two main leakage types:
    • Sub-threshold leakage: Current flows when gate "closed"
      • Threshold voltage lowers as gates shrink
      • Leakage rises ~10x for every 0.1V decrease
    • Gate tunneling: Electrons quantum tunnel through 1-2nm gate oxide
  • Problems emerged at 130nm node, major leap at 90nm node
  • Solutions: Gated VDD technique, FinFET introduction

FinFET Technology Impact

  • 3D gate covering channel on 3/4 sides
  • Better power leakage control and density
  • Old ITRS predicted 94% SRAM coverage by 2014 without changes
  • FinFETs avoided this scenario
  • Drawback: Harder to produce, higher defect risk

Process Variation and Yield Issues

  • Two yield types: Functional (works/doesn't work) and Parametric (performance variability)
  • Minute process variations impact final performance
  • Traced to lithography line control and roughness issues
  • Process defects don't scale down with size
  • Static Noise Margin (SNM): Measures cell resistance to electrical noise
  • Butterfly curve plots: Variations vs. SNM performance

Alternative SRAM Cell Designs

  • 5-transistor design: 15-20% less space, reduced stability
  • 10-transistor Schmitt trigger: 2x area, much more stable
  • 8-transistor: Good stability, bad power leakage
  • 9-transistor: Improved power/stability, more area/complexity
  • 6-transistor remains most used: Balance of simplicity, power, area, noise

Advanced Packaging Solutions

  • Companies adopting stacked SRAM arrangements
  • AMD 3D V-Cache: SRAM dies stacked on L3 cache
  • Advanced packaging pushed by SRAM scaling difficulties

Future SRAM Technology

  • TSMC hit limits with current FinFET architecture
  • N3B had yield issues, N3E SRAM density reversal indicates problems
  • Next: Gate-all-around FET/nanosheet designs
    • Gate wraps channel on all 4 sides
    • Samsung already manufacturing, Intel/TSMC/SMIC working on it
  • Fork sheets: Additional dielectric layer, better for continued scaling
  • SRAM scaling becoming "steadily losing game"
  • Need alternative solutions for fast cache memory

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