Crosspoint Array Negative Weight Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network technologies face challenges in efficiently computing negative weights, leading to increased area and energy consumption, as well as noise and variability, particularly in crosspoint device architectures used in artificial neural networks (ANNs).
Innovation Solution
The implementation of a crosspoint array with variable resistors and integrators, where each crosspoint stores a weight of the neural network, and a voltage source is applied to each conductive row, allowing for the determination of a reference voltage or current for each variable resistor, enabling accurate computation of negative weights while conserving area and maintaining circuit integrity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If crosspoint device architectures are used to implement neural networks with negative weights, then area consumption and energy consumption increase, but computing accuracy and reliability improve
Solution Approach 1:
The patent merges the reference voltage generation function directly into the crosspoint array structure by utilizing the conductive rows themselves as reference voltage lines. This integration eliminates the need for separate reference voltage circuits, thereby reducing area consumption while maintaining the ability to compute negative weights accurately through the established reference potential.
Solution Approach 2:
The conductive rows in the crosspoint array are designed to serve dual functions: they act as both signal transmission lines for neural network operations and as reference voltage lines for negative weight computation. This multi-functionality reduces the overall circuit complexity and area requirements while preserving computing accuracy.
2Reliability
If crosspoint device architectures are used to implement neural networks with negative weights, then area consumption and energy consumption increase, but processing reliability improves
Solution Approach 1:
The patent combines the reference voltage generation and signal transmission functions into the same conductive infrastructure. By using the existing conductive rows as reference voltage lines, the design eliminates redundant energy-consuming circuits while maintaining processing reliability through stable reference potentials for negative weight operations.
Solution Approach 2:
The crosspoint array structure provides its own reference voltage through the inherent properties of its conductive rows, eliminating the need for external reference voltage generation circuits. This self-service approach reduces energy consumption while maintaining the reliability needed for accurate negative weight computation.
3Ease of manufacture
If standard CMOS processing techniques are used for cross bar implementations, then manufacturing flexibility improves, but processing speed and efficiency decrease
Solution Approach 1:
The patent modifies the operational parameters of the crosspoint array by implementing specific voltage allocation schemes and integrator configurations that optimize processing speed. These parameter changes enable faster computation while remaining compatible with standard CMOS manufacturing processes, thus maintaining manufacturing flexibility.
4Measurement precision
If variable resistors and integrators are added to compute negative weights accurately, then computing accuracy improves, but device complexity increases
Solution Approach 1:
The patent integrates the reference voltage generation function into the existing conductive row structure, eliminating the need for separate reference voltage circuits. This merging reduces circuit complexity while maintaining the accuracy required for negative weight computation through the unified voltage reference system.
Solution Approach 2:
The conductive rows are designed to perform multiple functions simultaneously: signal transmission, reference voltage provision, and negative weight computation support. This multi-functionality reduces the overall device complexity by eliminating redundant components while preserving computing accuracy.
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 approach effectively computes negative weights with improved accuracy and reduced power consumption, enhancing the speed and efficiency of training ANN architectures while minimizing area penalties and noise.
Implementation Method 1
a set of conductive rows, a set of conductive columns intersecting the set of conductive rows to form a plurality of crosspoints
Implementation Method 2
a circuit element coupled to each of the plurality of crosspoints configured to store a weight of the neural network
Implementation Method 3
a first integrator attached at the end of at least one of the conductive column
Implementation Method 4
a first variable resistor attached to the integrator and the end of the at least one conductive column
Data Source
AI summary
Method, systems, crosspoint arrays, and systems for tuning a neural network. A crosspoint array includes: a set of conductive rows, a set of conductive columns intersecting the set of conductive rows to form a plurality of crosspoints, a circuit element coupled to each of the plurality of crosspoints configured to store a weight of the neural network, a voltage source associated with each conductive row, a first integrator attached at the end of at least one of the conductive column, and a first variable resistor attached to the integrator and the end of the at least one conductive column.