SOT Cell DNN Device for Energy-Efficient Inference

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current hardware implementations for deep neural networks (DNNs) during inference are limited in terms of energy efficiency, non-linearity, and density for AI applications.

Innovation Solution

A deep neural network device utilizing a plurality of spin-orbit torque (SOT) cells, where each SOT cell comprises a SOT layer, a ferromagnetic (FM) layer, and a controller to store weights of a neural network, enabling efficient matrix multiplication and activation functions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If traditional computing architecture is used for DNN inference, then data movement between memory and processor can be performed, but energy consumption increases and processing speed decreases

Engineering Contradiction:

Improveenergy consumptionVSAvoidarchitecture complexity

Core Design Contradiction:

Use of energy by moving objectVSDevice complexity

Solution Approach 1:

The patent merges memory and processing functions into a single compute-in-memory device. SOT cells perform both weight storage and matrix multiplication operations in-place, eliminating the need for separate memory and processor components. This integration directly reduces energy consumption by eliminating data movement while maintaining functional capability.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The SOT cell structure serves multiple functions: it stores neural network weights, performs matrix multiplication operations, and supports both training and inference modes. The same hardware infrastructure handles diverse computational tasks, reducing overall system complexity and energy usage compared to specialized separate components.

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Use of energy by moving object

If compute-in-memory hardware is implemented for DNNs, then energy consumption decreases and density increases, but non-linearity capability is limited

Engineering Contradiction:

Improveenergy consumptionVSAvoidnon-linearity capability

Core Design Contradiction:

Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The patent segments the neural network computation into distinct phases: linear matrix multiplication is performed by the SOT cell array, while non-linear activation functions are applied separately. This segmentation allows the hardware to excel at parallel multiplication while handling non-linearity through subsequent processing stages, maintaining both energy efficiency and computational versatility.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces intermediary processing stages between the SOT cell multiplication output and final results. These intermediaries handle activation functions and other non-linear operations, acting as a bridge between the efficient linear computation of SOT cells and the required non-linear transformations for complete DNN functionality.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Use of energy by moving object

If SOT cells are used for DNN inference, then energy consumption decreases and density increases, but current magnitude requirements increase

Engineering Contradiction:

Improveenergy consumptionVSAvoidcurrent magnitude

Core Design Contradiction:

Use of energy by moving objectVSPower

Solution Approach 1:

The patent optimizes SOT cell parameters including layer thickness, material composition, and geometric dimensions to reduce the current magnitude required for switching. By carefully tuning these parameters, the device achieves lower operational current requirements while maintaining the energy efficiency and high density benefits of compute-in-memory architecture.

Inventive Principle:

Principle #35Parameter changes

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 cell-based DNN device achieves lower energy consumption, higher non-linearity, and increased density for AI applications, improving the efficiency of DNN inference operations.

Implementation Method 1

each node of the n rows and m columns of nodes comprising a plurality of spin orbit torque (SOT) cells, each SOT cell comprising: at least one SOT layer, at least one ferromagnetic (FM) layer

Methodology Applied

Scientific EffectSpin-orbit torque:

Implementation Method 2

the FM layer may comprise two or more domains, two or more elliptical arms, or two or more states

Methodology Applied

Scientific EffectFerromagnetism: Ferromagnetism

Implementation Method 3

the FM layer being configured with a plurality of domain walls

Methodology Applied

Scientific EffectDomain walls:

Data Source

PatentUS20250077834A1In-Memory Deep Neural Network Device Using Spin Orbit Torque (SOT) With Multi-State Weight

Publication Date: 2025.03.06 WESTERN DIGITAL TECHNOLOGIES INC

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: at least one SOT layer, at least one ferromagnetic (FM) layer, and 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 SOT cell. The FM layer may comprise two or more domains, two or more elliptical arms, or two or more states.