SOT Magnetic Memory Array for Field-Free Vector-Matrix Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing magnetic memory devices face challenges in efficiently performing vector-matrix multiplication for artificial intelligence computation due to limitations in current technologies, particularly in terms of small current flow and neglecting spin-transfer torque effects.
Innovation Solution
A spin-orbit torque (SOT) magnetic memory device with an array of magnetic unit cells, each comprising a magnetic tunnel junction, is designed for compute-in-memory processing, utilizing an embedded electromagnet to generate a vertical magnetic field for programming magnetization states and enabling vector-matrix multiplication.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If voltage-controlled magnetic anisotropy (VCMA) is used to program magnetic unit cells, then the magnetoelectric unit cell can be toggled between parallel and anti-parallel states by applying voltage in one direction, but the current flow between the free layer and reference layer is very small, making spin-transfer torque (STT) effects negligible and limiting computational efficiency
Solution Approach 1:
The patent introduces a spin-orbit torque (SOT) layer as an intermediary component between the electrical signal and the magnetic tunnel junction. This SOT layer converts the applied voltage into a spin current that acts on the magnetic moments, enabling efficient magnetization switching. The SOT layer serves as a mediator that transforms electrical energy into magnetic effects through spin-orbit coupling, resolving the contradiction between ease of voltage-based operation and computational efficiency.
Solution Approach 2:
The patent changes the operational parameter from direct voltage application across the magnetic tunnel junction to voltage application across a spin-orbit coupling layer. This parameter change enables spin-transfer torque effects to become significant by generating a spin-polarized current that can efficiently switch magnetization states, thereby improving computational efficiency while maintaining voltage-controlled operation.
2Device complexity
If conventional magnetic memory devices are used, then the structure is relatively simple, but they cannot efficiently perform vector-matrix multiplication for artificial intelligence computation
Solution Approach 1:
The patent designs the magnetic memory device to serve multiple functions: it can store data in conventional modes and simultaneously perform vector-matrix multiplication operations. By integrating spin-orbit torque layers and configuring the magnetic tunnel junctions with appropriate magnetic anisotropy, the device can execute neural network computations directly in memory, achieving multi-functionality that resolves the contradiction between structural simplicity and computational capability.
Solution Approach 2:
The patent merges the memory storage function with the computational function by combining magnetic memory cells with spin-orbit torque layers in a unified structure. This merging allows the same physical structure to both store information and perform computations, eliminating the need for separate memory and processing units, thereby achieving efficient AI computation without significantly increasing overall device complexity.
3Productivity
If spin-orbit torque (SOT) magnetic memory device is used with embedded electromagnet for vector-matrix multiplication, then computational efficiency is enhanced, but the device complexity increases due to additional components and configuration requirements
Solution Approach 1:
The patent segments the magnetic memory device into distinct functional layers: spin-orbit torque layers, magnetic tunnel junctions with perpendicular magnetic anisotropy, and electrode structures. Each segment performs a specific function, allowing for modular design and fabrication. This segmentation enables the complex computational functionality to be achieved through well-defined, separately optimizable components, managing device complexity through systematic division of functions.
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
The SOT magnetic memory device enhances computational efficiency by allowing simultaneous programming of multiple cells and performing vector-matrix operations, thereby improving artificial intelligence computations.
Implementation Method 1
a spin-orbit torque (SOT) layer, e.g., a metal layer exhibiting the spin Hall effect
Implementation Method 2
Voltage-controlled magnetic anisotropy (VCMA) refers to magnetic anisotropy that increases or decreases with application of an electric field across a magnetic tunnel junction
Implementation Method 3
the current is typically so small that spin-transfer torque (STT) effects can be ignored
Data Source
AI summary
A device includes an array of magnetic unit cells located over a substrate, where each of the magnetic unit cells includes a magnetic tunnel junction, first nonmagnetic, electrically conductive lines electrically contacting respective row magnetic tunnel junctions, second nonmagnetic, electrically conductive lines contacting a respective column of magnetic tunnel junctions, and a soft magnetic material layer.