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
- Sub-threshold leakage: Current flows when gate "closed"
- 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
