3D Chiplet Memory-Logic Stacking for Dense CIM Interconnects

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current compute-in-memory (CIM) and compute-near-memory (CNM) architectures face challenges in integrating with existing processors due to device node mismatches in two-dimensional (2D) SoC designs and interconnect density requirements in three-dimensional (3D) stacked chips, limiting their feasibility for energy-efficient computing in edge accelerators.

Innovation Solution

The proposed solution involves a 3D chiplet architecture with wafer-on-wafer stacked memory and control tiers, utilizing direct bonds and oxide bonds to simplify interconnects, and through-silicon vias (TSVs) for data links between logic and memory tiers, enabling high-density, low-power CIM chiplets compatible with CMOS technology for integration with CPUs, GPUs, and AI accelerators.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If 2D SoC design is used for CIM architecture, then device node mismatch challenges occur between NVM function block and control logic block, but integration with existing processors is simplified

Engineering Contradiction:
Improveintegration compatibilityVSAvoiddevice node matching
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The patent transitions from 2D SoC design to 3D stacked chiplet architecture, stacking memory die and logic die vertically to achieve both integration compatibility and device node matching. The 3D configuration allows independent optimization of memory and logic layers while maintaining interconnect density through vertical stacking with microbumps and TSVs.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Adaptability or versatility

If 3D stacked chips with microbump technology are used, then interconnect density requirements for CIM architecture partitioning and re-integrating are not fulfilled, but stacking of memory and logic tiers is achieved

Engineering Contradiction:
ImproveCIM architecture partitioningVSAvoidinterconnect density
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent segments the CIM architecture into separate memory die and logic die that are stacked vertically. This segmentation allows independent optimization of each layer while achieving high interconnect density through the stacking interface using microbumps and through-silicon vias, enabling efficient partitioning and re-integration of CIM components.

Inventive Principle:
Principle #1Segmentation

3Adaptability or versatility

If monolithic 3DIC is used for memory architecture, then feasibility for CIM and CNM implementations is limited, but three-dimensional integration is achieved

Engineering Contradiction:
Improvememory architecture implementationVSAvoidmonolithic integration complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent divides the monolithic 3DIC into separate stackable chiplets including memory die and logic die. This segmentation reduces the complexity of monolithic integration while maintaining three-dimensional integration benefits, allowing flexible CIM and CNM implementations through heterogeneous stacking of specialized functional blocks.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS20230420437A1Three dimensional (3D) chiplet and methods for forming the same
Publication Date: 2023.12.28 TAIWAN SEMICONDUCTOR MANUFACTURING CO LTD
  • US20230420437A1 patent drawing
  • US20230420437A1 patent drawing
  • US20230420437A1 patent drawing

AI summary

A semiconductor structure, includes a logic die, a memory die stack bonded to the logic die by a first oxide bond, and including a first pair of memory dies bonded together by a first direct bond, and a first through silicon via (TSV) in the logic die and extending across the first oxide bond and electrically connecting the logic die to the first pair of memory dies.