Programmable Neuron Circuits for VMM Current-Adaptive Power Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing artificial neural networks face challenges in high-performance information processing due to inadequate hardware technology, particularly in energy efficiency and flexibility of summer and activation function circuits, which are not adaptable to varying current outputs from vector-by-matrix multiplication (VMM) arrays.

Innovation Solution

The integration of CMOS technology with non-volatile memory arrays and the development of adjustable summer and activation function circuits that can be configured to optimize power consumption based on the total current received from VMM arrays, utilizing variable resistors and control circuits to adjust resistance and minimize power consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If prior art summer and activation circuits are used, then the neural network can operate, but the circuits cannot be configured for each particular VMM array and cannot optimize power consumption

Engineering Contradiction:
Improveadaptability to varying current outputsVSAvoidcircuit configuration complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements dynamic configuration of the summer and activation function circuits through programmable control. The summer circuit includes programmable gain amplifiers that can be configured via control signals to match different VMM array current outputs. The activation function circuit uses programmable threshold voltages and gain factors that can be adjusted based on the specific VMM array being used, enabling adaptability without permanent hardware changes.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes key circuit parameters such as gain factors, threshold voltages, and reference currents to optimize performance for each VMM array. The summer circuit allows programming of gain parameters to match the total current output capability of different VMM arrays. The activation function circuit programs threshold and slope parameters to adapt to varying input ranges, resolving the contradiction between adaptability and complexity through software-controlled parameter adjustment.

Inventive Principle:
Principle #35Parameter changes

2Use of energy by moving object

If fixed summer and activation circuits are used, then device complexity is reduced, but power consumption cannot be optimized based on VMM current output

Engineering Contradiction:
Improvepower consumptionVSAvoidcircuit adjustability
Core Design Contradiction:
Use of energy by moving objectVSDevice complexity

Solution Approach 1:

The patent employs dynamic power management through programmable circuit elements. The summer circuit includes programmable gain amplifiers with adjustable bias currents that can be scaled according to the VMM array's current output level. The activation function circuit uses programmable threshold voltages and gain factors that enable operation at lower power levels when the VMM array produces smaller current signals, optimizing energy consumption while maintaining adaptability through control logic.

Inventive Principle:
Principle #15Dynamics

3Use of energy by moving object

If non-volatile memory arrays are used as synapses, then energy efficiency is improved, but precise programming of floating gate charge is required

Engineering Contradiction:
Improveenergy efficiencyVSAvoidfloating gate charge precision
Core Design Contradiction:
Use of energy by moving objectVSManufacturing precision

Solution Approach 1:

The patent implements feedback mechanisms during the programming process of non-volatile memory cells. The system uses read-verify-program cycles where the floating gate charge is programmed, then read back to verify the weight value, and adjusted if necessary. This feedback loop ensures precise weight storage in the non-volatile memory array while maintaining energy efficiency, as the feedback control compensates for variations in programming precision.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3735656B1Programmable neuron for analog non-volatile memory in deep learning artificial neural network
Publication Date: 2023.12.27 SILICON STORAGE TECHNOLOGY INC
  • EP3735656B1 patent drawingFigure 1
  • EP3735656B1 patent drawingFigure 2
  • EP3735656B1 patent drawingFigure 3

AI summary

Numerous embodiments for processing the current output of a vector-by-matrix multiplication (VMM) array in an artificial neural network are disclosed. The embodiments comprise a summer circuit and an activation function circuit. The summer circuit and/or the activation function circuit comprise circuit elements that can be adjusted in response to the total possible current received from the VMM to optimize power consumption.