Dual Spin Orbit Torque Device for Energy Efficient DNN Compute
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current hardware implementations for deep neural networks (DNNs) are limited in terms of energy efficiency, speed, and integration density, primarily due to reliance on traditional computing architectures with discrete memory and processor components.
Innovation Solution
The use of spin orbital-spin orbital (SO-SO) devices, which comprise two spin orbit torque (SOT) layers and a ferromagnetic layer, configured to perform various functions of a neural node in a DNN, including storing weights, multiplying inputs and weights, summing results, and performing activation functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional computing architecture with discrete memory and processor components is used, then the DNN can perform matrix multiplication and activation functions, but energy consumption increases and speed decreases
Solution Approach 1:
The patent combines memory and processing functions into a single compute-in-memory device. The crossbar array structure allows weight storage in resistive memory elements while simultaneously performing multiplication and accumulation operations through analog voltage division, eliminating the need for separate memory read and processing steps.
Solution Approach 2:
The patent replaces digital binary computation with analog continuous voltage representation. Neural network weights and activations are represented as continuous voltage levels rather than discrete binary values, enabling parallel analog computation that consumes less energy and operates faster than sequential digital processing.
2Ease of manufacture
If traditional computing architecture is used, then the DNN can be implemented with standard components, but integration density decreases
Solution Approach 1:
The patent transitions from planar 2D circuit layouts to three-dimensional stacked architectures. Multiple computational layers are stacked vertically with through-silicon vias connecting them, dramatically increasing the number of neural network layers that can be integrated in a given footprint while maintaining compatibility with standard semiconductor manufacturing processes.
Solution Approach 2:
The patent designs a universal compute-in-memory cell structure that can perform multiple neural network operations (multiplication, accumulation, activation) within a single integrated device. The same crossbar array infrastructure supports both weight storage and computational functions, reducing the number of separate components needed compared to traditional architectures.
3Use of energy by moving object
If compute-in-memory hardware is used, then energy consumption decreases and integration density increases, but the implementation becomes limited in functionality
Solution Approach 1:
The patent implements dynamically reconfigurable compute-in-memory arrays where the functional configuration can be changed through programmable control signals. The same hardware infrastructure can be dynamically switched between different computational modes (e.g., MAC operations, activation functions, transpose operations) to support various neural network architectures and operations without requiring dedicated hardware for each function.
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 SO-SO device efficiently provides spin-to-charge and charge-to-spin mechanisms, enabling flexible configuration to perform neural node functions while reducing energy consumption and improving integration density.
Implementation Method 1
two spin orbit torque (SOT) layers, a first spin orbit torque (SOT1) layer, a second spin orbit torque (SOT2) layer
Implementation Method 2
The SO-SO device thus efficiently provides spin-to-charge and charge-to-spin mechanisms in the same device
Implementation Method 3
The SO-SO device thus efficiently provides spin-to-charge and charge-to-spin mechanisms in the same device
Implementation Method 4
a ferromagnetic layer disposed between the SOT1 and SOT2 layer
Data Source
AI summary
The present disclosure generally relates to a deep neural network (DNN) device utilizing spin orbital-spin orbital (SO-SO) devices. The SO-SO devices each includes two SOT layers, a first spin orbit torque (SOT1) layer, a second spin orbit torque (SOT2) layer, and a ferromagnetic layer disposed between the SOT1 and SOT2 layer. Each SO-SO device further comprises three terminals, one per each SOT layer, for in plane current flow to or from the respective SOT layer, and one for perpendicular current flow through multiple layers, or the overall stack, of the SO-SO device. The SO-SO device thus efficiently provides spin-to-charge and charge-to-spin mechanisms in the same device, and can be flexibility configured to perform various functions of a neural node of a DNN. These functions include storing programmed weights, multiplying inputs and weights and summing such multiplication results, and performing an activation function to determine a neural node output.