Multi-State SOT Weight Cells for Low-Energy In-Memory DNN Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current hardware implementations of deep neural networks (DNNs) for inference are limited in terms of energy efficiency, memory consumption, and speed due to reliance on traditional Von Neumann architecture, which requires significant data movement between memory and processing units.
Innovation Solution
A DNN device utilizing spin-orbit torque (SOT) cells with ferromagnetic layers and magnetic domains, where weights are stored in oxide layers with varying magnetic anisotropy, enabling efficient matrix multiplication and activation functions within the memory array.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional Von Neumann architecture is used for DNN inference, then data movement between memory and processing units is enabled, but energy consumption increases and computational speed decreases
Solution Approach 1:
The patent merges memory and processing functions by implementing matrix multiplication capabilities directly within the memory array. The cross-point memory architecture allows weights stored in memory cells to be multiplied with input signals during inference, eliminating the need to move data between separate memory and processing units, thus reducing energy consumption while maintaining functional capability
2Speed
If traditional Von Neumann architecture is used for DNN inference, then discrete memory and processor components are used, but computational speed decreases due to data movement requirements
Solution Approach 1:
The patent combines memory storage and computational processing into a single integrated structure. The cross-point memory array performs matrix multiplication in-place during inference operations, eliminating data transfer between separate memory and processor components, thereby increasing computational speed while reducing energy loss from data movement
3Productivity
If compute-in-memory hardware implementation is used, then energy consumption decreases and computational density increases, but implementation complexity increases
Solution Approach 1:
The patent implements a universal cross-point memory architecture that can store weights and perform matrix multiplication operations simultaneously. The same memory structure serves both storage and computation functions, increasing computational density while managing implementation complexity through a unified design approach rather than separate specialized components
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 reduces energy consumption and increases computational density by performing matrix-vector multiplication directly in memory, enhancing the efficiency of DNN inference operations.
Implementation Method 1
Each FM domain is disposed in contact with an oxide layer, which can induce low magnetic anisotropy (Ku) inside the FM layer (a low Ku oxide layer), and an oxide that can induce high Ku inside the FM layer (a high Ku oxide layer)
Implementation Method 2
each node of the n rows and m columns of nodes comprising a plurality of spin orbit torque (SOT) cells, each SOT cell comprising: a SOT layer, and a ferromagnetic (FM) layer
Data Source
AI summary
The present disclosure is generally related to a deep neural network (DNN) device comprising a plurality of spin-orbit torque (SOT) cells. The DNN device comprises an array comprising n rows and m columns of nodes, each row of nodes coupled to one of n first conductive lines, each column of nodes coupled to one of m second conductive lines, each node of the n rows and m columns of nodes comprising a plurality of SOT cells, each SOT cell comprising: a SOT layer, and a ferromagnetic (FM) layer comprising two or more magnetic domains. Each domain is disposed in contact with a low magnetic anisotropy (Ku) oxide layer, and a high Ku oxide layers. The DNN device further comprises a controller configured to store at least one corresponding weight of an n×m array of weights of a neural network in each of the domains.