Programmable Analog Neuron Circuits for VMM Current-Aware Power Tuning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing artificial neural networks face challenges in optimizing power consumption due to inflexible summer and activation circuits that cannot be adjusted to account for the varying maximum current output of vector-by-matrix multiplication (VMM) arrays.

Innovation Solution

The development of adjustable summer and activation function circuits that can be configured to optimize power consumption by responding to the total possible current received from the VMM, utilizing CMOS technology and non-volatile memory arrays to enable precise tuning of synapse weights and efficient energy use.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If fixed summer and activation circuits are used in artificial neural networks, then circuit simplicity is maintained, but power consumption cannot be optimized for varying VMM current outputs

Engineering Contradiction:

Improvepower consumptionVSAvoidcircuit adjustability

Core Design Contradiction:

Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamically adjustable summer and activation circuits that can adapt their parameters based on the actual current output characteristics of the VMM array. The summer circuit includes adjustable gain control, and the activation function circuit includes可调 threshold and slope parameters, allowing the system to optimize power consumption for different operating conditions while maintaining adaptability.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The invention changes key circuit parameters (gain, threshold, activation slope) from fixed values to adjustable values that can be programmed based on the VMM array's current output range. This allows the neural network to optimize power consumption by matching circuit parameters to the actual computational workload and current characteristics.

Inventive Principle:

Principle #35Parameter changes

2Use of energy by moving object

If non-volatile memory arrays are used for synapses, then energy efficiency is improved through analog computation, 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 incorporates feedback mechanisms during the programming process of non-volatile memory cells. The system monitors the current output from the VMM array and uses this information to adjust programming parameters, ensuring that the floating gate charge is programmed to the precise desired level. This feedback loop compensates for process variations and achieves accurate weight storage.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The invention performs preliminary characterization of the VMM array's current output range before configuring the summer and activation circuits. This preliminary action allows the system to pre-calculate optimal parameter settings and programming conditions, ensuring that subsequent operations achieve both energy efficiency and precision without requiring excessive programming iterations.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11354562B2Programmable neuron for analog non-volatile memory in deep learning artificial neural network

Publication Date: 2022.06.07 SILICON STORAGE TECHNOLOGY INC

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.