3D Neuromorphic Memory Using Ternary Data Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neuromorphic devices lack efficient methods for storing and processing ternary data, which is essential for advanced neural network operations, and there is a need for improved data storage and reading techniques in three-dimensional neuromorphic systems.
Innovation Solution
A three-dimensional neuromorphic device is developed using phase change elements and ovonic threshold switches, with a controller that converts binary data into ternary data and stores it using a memory device with separate arrays for synaptic and weight values, enabling efficient data storage and reading through a tile-based structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If binary data storage methods are used in conventional memory devices, then data storage is simple and reliable, but ternary data processing capability is insufficient for advanced neural network operations
Solution Approach 1:
The memory device is divided into two separate memory cell arrays: a first memory cell array for storing data bits and a second memory cell array for storing mask bits. This segmentation allows the system to handle ternary data (0, 1, 2) by combining binary data from both arrays, achieving ternary processing capability while keeping each individual array structurally simple and manageable.
Solution Approach 2:
The patent transitions from conventional single-dimension binary storage to a two-dimensional architecture where data bits and mask bits are stored in separate arrays that can be independently accessed. This dimensional expansion enables ternary data representation by combining results from both arrays, effectively adding a new dimension to data processing capability.
2Productivity
If separate memory cell arrays are used for data bits and mask bits, then ternary data storage efficiency is improved, but device structure and control complexity increase
Solution Approach 1:
Both the first and second memory cell arrays use the same basic memory cell structure comprising a switching element and a resistive element. This universal design allows both arrays to perform identical storage functions while being controlled differently to achieve ternary data processing, improving storage efficiency without requiring fundamentally different structures for each array.
Solution Approach 2:
The patent pre-organizes data into data bits and mask bits before storage, with data bits stored in the first array and mask bits stored in the second array. This preliminary organization enables efficient ternary data retrieval and processing by allowing independent access and combination of the two bit types, improving productivity while managing structural complexity through systematic control.
3Speed
If conventional reading methods are used, then reading process is simple, but reading speed and efficiency for ternary data are insufficient
Solution Approach 1:
The patent enables simultaneous or sequential reading operations from both the first memory cell array (data bits) and the second memory cell array (mask bits). By allowing continuous reading actions from both arrays and immediately combining the results, the system achieves high-speed ternary data retrieval without requiring complex intermediate processing steps, maintaining reading efficiency while supporting enhanced functionality.
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 solution enables efficient storage and processing of ternary data, facilitating advanced neural network operations by improving data storage and reading techniques in three-dimensional neuromorphic systems, enhancing the performance of neuromorphic devices.
Implementation Method 1
a first memory cell of the first memory cells includes a first switching element and a first resistive element
Implementation Method 2
a phase change element and an ovonic threshold switch
Data Source
AI summary
A neuromorphic device, including a controller configured to generate ternary data by converting each bit of binary data into a ternary bit; and a memory device configured to store the ternary data, wherein the memory device includes: a first memory cell array including first memory cells formed between lower word lines and bit lines, wherein a first memory cell of the first memory cells includes a first switching element and a first resistive element; and a second memory cell array including second memory cells formed between upper word lines and the bit lines, wherein a second memory cell of the second memory cells includes a second switching element and a second resistive element, and wherein each bit of the ternary data is identified by a combination of a data bit stored in the first memory cells and a mask bit stored in the second memory cells.