Analog Memory Tuning With Light and Heat for Scalable In-Memory Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for tuning analog memory elements in neomorphic computing systems, such as electrochemical tuning, introduce complexity, limit scalability, and impose thermal management constraints, hindering efficient AI computations.
Innovation Solution
Tuning analog memory elements using application of light and/or heat, which allows for reversible resistance changes in transition metal oxides like TiOx, WOx, VOx, NbOx, or silicon oxide, enabling scalable and thermally efficient in-memory computing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If electrochemical methods are used to tune analog memory elements, then conductance tuning is achieved, but device complexity increases and scalability is limited
Solution Approach 1:
The patent replaces electrochemical tuning methods with optical and thermal methods. Light and heat are applied directly to the analog memory elements to induce resistance changes, eliminating the need for complex electrochemical structures and processes on the microchip. This substitution reduces device complexity while maintaining tuning precision.
Solution Approach 2:
The patent changes the tuning mechanism from electrochemical to optical/thermal parameters. By controlling light wavelength, intensity, and duration, or temperature and exposure time, the resistance of analog memory elements is tuned. This parameter-based approach simplifies the device structure while enabling precise conductance control.
2Manufacturing precision
If electrochemical methods are used to tune analog memory elements, then conductance tuning is achieved, but the analog memory array scale is limited
Solution Approach 1:
The patent replaces electrochemical methods with optical and thermal methods that can be applied uniformly across large arrays of analog memory elements. Light and heat can be distributed over large areas simultaneously, enabling scaling to thousands or millions of memory elements without the complexity constraints of electrochemical approaches.
Solution Approach 2:
The optical and thermal tuning methods are universally applicable to all analog memory elements in the array, regardless of position or size. This universal approach enables uniform tuning across large-scale arrays, making it possible to scale the quantity of memory elements while maintaining consistent tuning precision.
3Manufacturing precision
If electrochemical methods are used to tune analog memory elements, then conductance tuning is achieved, but thermal management constraints are imposed
Solution Approach 1:
The patent substitutes electrochemical methods with optical and thermal methods that do not generate significant heat. Light-induced resistance changes and controlled thermal exposure allow for precise tuning without the thermal management constraints associated with electrochemical processes, enabling operation at lower temperatures.
4Productivity
If conventional CMOS hardware is used for AI computations, then processing capability is achieved, but power consumption is high
Solution Approach 1:
The patent implements in-memory computing where analog memory elements perform computational tasks directly within the memory array itself, eliminating the need to transfer data between memory and processing units. This self-service approach reduces energy consumption by avoiding repeated data transfers while maintaining high computation speed for AI tasks.
Solution Approach 2:
The patent utilizes phase changes in analog memory elements (such as transition metal oxides undergoing phase transitions between different resistance states) to enable computational operations. These phase transitions allow for low-power switching and processing, significantly reducing the power consumption compared to conventional CMOS hardware while maintaining AI computation capabilities.
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 method enables scalable, low-complexity, and thermally efficient tuning of analog memory elements, suitable for complex AI computations like matrix vector operations and deep neural networks, without the limitations of electrochemical methods.
Implementation Method 1
application of light and/or heat promotes reactions required for insertion of ions into the analog memory element and/or removal of oxygen vacancies from the analog memory element, thereby tuning the underlying electronic structure of the analog memory element (and thus tuning the resistance of the analog memory element)
Implementation Method 2
application of heat and/or light to analog memory elements can initiate phase changes in the analog memory elements, and therefore the resistance of the material can be altered by way of phase change (where oxides are an example phase change material)
Data Source
AI summary
A method is disclosed herein which includes obtaining an array of analog memory elements. It also includes programming an analog memory element included in the array, where prior to being programmed the analog memory element has a first value for an electrical property, and where it is programmed to cause the analog memory element to perform a computation included in a series of computations performed by the array. Programming the analog memory element includes applying light or heat to the analog memory element, where a value of the electrical property is changed from the first value to a second value based upon application of light or heat to the analog memory element, and further where upon the value of the electrical property being changed from the first value to the second value, the analog memory element is configured to perform the computation responsive to receipt of an input.