3D Stacked Neural Network Accelerator Tiles via Inductive Coupling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As neural network models increase in size and complexity, they require more computational resources, leading to a need for large-scale hardware neural network accelerators that can efficiently handle these demands while maintaining scalability, cost-effectiveness, and bandwidth.
Innovation Solution
A three-dimensionally stacked neural network accelerator architecture utilizing vertically aligned tiles with inductive coupling through ThruChip Interface (TCI) technology, forming a static interconnect system for uninterrupted data flow and digital logic interconnections, which enhances on-chip memory capacity, scalability, and reduces costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If three-dimensional stacking is used to increase computational capacity, then the neural network accelerator can handle larger models with increased memory capacity, but the wiring complexity and manufacturing difficulty increase
Solution Approach 1:
The patent transitions from two-dimensional planar arrangements to three-dimensional stacked architecture, enabling vertical integration of multiple neural network accelerator tiles. This dimensional change allows increased memory capacity and computational resources without proportionally increasing wiring complexity, as connections are established through vertical stacking rather than extensive lateral routing.
Solution Approach 2:
The system divides the neural network accelerator into multiple independent tiles that can be stacked vertically. Each tile is a self-contained unit with its own processing elements and memory, allowing modular assembly and reducing the complexity of interconnections compared to a monolithic design.
2Productivity
If three-dimensional stacking with wireless communication is implemented, then bandwidth and data flow continuity are improved, but manufacturing precision requirements increase
Solution Approach 1:
The patent introduces transmission coils as intermediary components that enable wireless communication between stacked tiles. These coils act as mediators that transfer data and power across the vertical gap between tiles without requiring direct physical contact, thereby reducing the precision requirements for alignment while maintaining high bandwidth communication.
Solution Approach 2:
The system replaces mechanical/electrical connection systems (such as physical contacts or through-silicon vias) with wireless inductive coupling through transmission coils. This substitution eliminates the need for precise mechanical alignment and physical contact points, reducing manufacturing precision requirements while enabling continuous data flow.
3Adaptability or versatility
If vertical stacking with transmission coils is used, then scalability and compactness are improved, but cost of implementation increases
Solution Approach 1:
The transmission coils serve multiple functions simultaneously: they enable wireless data communication, provide power transfer between tiles, and facilitate alignment during assembly. This multi-functionality reduces the need for separate components and processes, thereby lowering overall manufacturing costs despite the advanced technology involved.
Solution Approach 2:
The patent combines communication and power transfer functions into a single wireless inductive coupling system using transmission coils. By merging these functions rather than implementing separate wired connections for data and power, the system reduces the number of manufacturing steps and components required, improving scalability while controlling costs.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach enables increased on-chip memory, lower costs, higher bandwidth, and greater compactness, facilitating efficient neural network computations by allowing larger models to be processed with improved scalability and reduced wiring complexity.
Implementation Method 1
the first transmission coil is configured to establish wireless communication with the second transmission coil via inductive coupling
Data Source
Figure 1A~1C
Figure 2
Figure 3
AI summary
A three dimensional neural network accelerator that includes a first neural network accelerator tile that includes a first transmission coil, and a second neural network accelerator tile that includes a second transmission coil, wherein the first neural network accelerator tile is adjacent to and aligned vertically with the second neural network accelerator tile, and wherein the first transmission coil is configured to wirelessly communicate with the second transmission coil via inductive coupling.