AI summary by TwinMind
Stacking Dies For Performance and Profit
Asianometry
14m 44s
Introduction & Definitions
- Advanced packaging technologies emerging as Moore's Law ends
- Die stacking at heart of next-generation performance gains
- Multi-chip package/system-in-package solutions contain multiple dies in single package
- 2.5D integration: co-packages dissimilar parts (logic, memory) side-by-side on connective substrate
- 3D integration: co-packages dissimilar parts in vertical stack using various connection technologies
- No formal definitions exist for 2.5D vs 3D classifications
Motivation for Advanced Packaging
- Moore's Law dead - can still cram more components but costs onerous for all but richest customers
- Industry coalesced toward two directions:
- Add accelerator chips next to CPUs for specific algorithms
- Optimize costs via chiplet approach - partition monolithic chip, scale only critical parts
- 2.5D and 3D solutions enable both directions
2.5D Integration History & Evolution
- Side-by-side die stacks date to 1960s hybrid circuits
- Combined devices like ICs and RF/microwave units on substrate
- Ceramic substrates with bonded wires
- Welded lids for protection
- Multi-chip modules (MCMs) emerged 1980s
- Direct descendants of hybrid circuits
- More complex structure with 2+ ICs electrically connected to common substrate
- Interconnect substrates (high density interconnects) designed/manufactured side-by-side
- Original goal: build faster systems, reduce packaging delays
- Delays went from small percentage to nearly 100% of cycle time
- Mainframe makers like Unisys tried late 1970s for high performance computing
- Ultimately easier to let Moore's Law increase silicon density
- MCMs fell by wayside as single die eliminated packaging delay
Modern 2.5D Revival
- 2009: Dr. Tong Ho Ming (ASE Group head of R&D) coined "2.5D" phrase
- Declared industry might need intermediate step to reach true 3D integration
- 2012: TSMC announced 2.5D chip-on-wafer-on-substrate technology
- Partnered with big EDA companies
- UMC and GlobalFoundries announced solutions around same time
- Actual work handled by OSAT partners
2.5D Key Technology: Through Silicon Interposers
- Uses through silicon interposer sitting on substrate
- Features through silicon vias (TSVs)
- TSVs: vertical interconnects - wrapped conductor passing through die's silicon substrate
- Concept invented by William Shockley 1958 (called "deep pits")
- Made using deep reactive ion etching - corrosive gases to rapidly etch silicon
- Originally developed for MEMS industry
Early 2.5D Products
- First 2.5D packages emerged 2010s:
- AMD Radeon Fury GPUs
- Xilinx Vertex FPGAs
- Vendors chose TSV implementation for less execution risk
2.5D Advantages
- Less space/thermal limitations on dies/packages that can be stacked
- More physical space for in-package memory capacity
- Critical for memory-intensive AI tasks facing "memory wall"
- Memory wall: efficiency/speed of data transfer between chip logic and memory
- No significant re-engineering of chips required - uses conventional dies
- With true 3D, dies need TSVs drilled through them, must be significantly thinner
2.5D Disadvantages
- Additional cost of producing silicon interposer
- Another point of failure
- AMD and ASE Group had to overcome warpage in silicon interposer
- Industry exploring new 2.5D generations removing silicon interposer entirely
- Chips sit directly on substrate
Chiplet Revolution
- Industry focus on splitting monolithic die into smaller chiplets
- Better yield, lower costs without compromising performance
- Universal Chiplet Interconnect Express (UCIe) standard introduced this year
- Open specification for chiplets
- Enables mixing/matching dies from different manufacturers
- Backed by AMD, Intel, ARM, Nvidia, ASE Group, Samsung, TSMC
- US-China chip sanctions may affect chiplet interconnect standard
- Potential problems transferring technology to Chinese firms like Alibaba
- UCIe 1.0 defined for 2D and 2.5D packaging - indicates 2.5D fully matured
3D Stacking Historical Context
- Vertical stacking predates transistor era
- 1950s: US Navy "Tinker Toy" circuits - large stacked ceramic substrates for vacuum tubes
- 1958: RCA "Micro Modules" project for US Army
- Army wanted rugged, mini-sized electronics
- $5 million funding
- 193 square millimeter ceramic substrates, vertically stackable
- Failed due to high costs, integrated circuits offered better prospects
- Soviets cloned anyway
Package-on-Package Evolution
- Mid-1980s: vendors needed better PCB real estate usage
- Consumer gadgets getting smaller drove stacked packages
- Smartphone vendors adopted late 2000s
- iPhone 5S example (September 2013):
- LPDDR3 memory wire bonded to package
- Memory package stacked on Apple A7 SoC
- A7 packaged using flip chip (upside down with bumps/balls vs bond wires)
- Limitations: package-within-package adds weight, can't stack many
- Suboptimal interconnect - signals travel package-to-package
Die-to-Die 3D Integration
- Vertical stacks of unpackaged dice
- Interconnections via traditional die-to-die wire bonding or TSVs
- Traditional wire bonding popular today:
- Piggyback style with bond wire connections
- Available since 1990s
- Very small, thin electronics (hearing aids)
- TSV-connected die stacking seen as more future potential:
- Stack memory dies directly on processor die
- Shortest possible interconnect route
- Potentially thousands of routes
- "Holy Grail" of integration
3D Implementation Examples
- First TSV-connected non-memory die stacks from CMOS image sensor world
- Memory industry went 3D early 2010s
- CMOS image sensors made of active pixels:
- Photodiode for light collection
- Extra transistors for complementary functions
- Fill factor concern: percentage of pixel space for light collection
- 3D stacking lets photodiode get bigger, approaching 100% fill factor
- Sony announced breakthrough 2005, started production 2012
3D Manufacturing Challenges
- TSV formation while avoiding failure modes:
- Very small: 2-10 micrometers wide, 40-200 micrometers long
- Multi-step process: etching, lining, filling, revealing
- Cooling requirements:
- Lower memory temperatures for long-term reliability
- AMD uses blank silicon chiplets for heat conduction
- Reliability concerns:
- Vendors worried about defects during manufacturing stages
- New systems needed for testing/repairing failures
Current 3D Products
- AMD V3D processors
- Graphcore Bow AI accelerators (using TSMC advanced packaging)
- Intel Ponte Vecchio processor (using Intel Faveros technology for stacking chiplets/tiles)
TSMC N3 Process & Future
- December 29, 2022: TSMC ceremony in Tainan marking N3 (3nm) volume production
- Fulfilled promise for second half 2022
- N3 development troubled
- 68th annual IEEE International Electron Devices Meeting findings:
- N3B (first 3nm process): minimal SRAM bit cell size shrinking (~5% or 0.95X scaling)
- N3E (next version): no SRAM shrinkage at all
- SRAM shrinkage issues require deeper chiplet-style integration
- Need improved system-level performance, especially for memory needs
- TSMC's 3D die stacking offerings hitting market at right time
