Analog Neural Memory Output Blocks for Precise VMM Weight Verification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing artificial neural networks face challenges in implementing effective input and output mechanisms for vector-by-matrix multiplication (VMM) arrays, particularly in accurately reading and verifying values stored in non-volatile memory cells, which are crucial for precise weight tuning and efficient energy usage.

Innovation Solution

The development of output blocks that include current-to-voltage converters and analog-to-digital converters to generate output bits from non-volatile memory cells, enabling precise programming, reading, and verification of memory cell values, utilizing CMOS technology and non-volatile memory arrays to enhance the efficiency of neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If non-volatile memory arrays are used as synapses in analog neural networks, then energy efficiency is improved, but the ability to accurately read and verify stored weight values deteriorates

Engineering Contradiction:

Improveenergy efficiencyVSAvoidreading accuracy of memory cell values

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent introduces an output block as an intermediary component between the non-volatile memory array and the digital processing system. This output block includes current-to-voltage converters and analog-to-digital converters that mediate the signal transformation process, enabling accurate reading of analog weight values stored in the memory cells while maintaining the energy efficiency benefits of non-volatile memory usage.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces direct digital reading mechanisms with analog signal conversion mechanisms. Instead of attempting to directly read digital values from the memory cells, the system uses current-to-voltage conversion followed by analog-to-digital conversion, substituting a multi-stage analog processing approach for a simpler direct digital read approach, thereby improving measurement precision.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Use of energy by moving object

If analog computation is used in neural networks, then energy consumption is reduced, but precision in representing weight values deteriorates

Engineering Contradiction:

Improveenergy consumptionVSAvoidprecision of weight values

Core Design Contradiction:

Use of energy by moving objectVSManufacturing precision

Solution Approach 1:

The patent changes the parameter representation by storing weight values as analog quantities (currents or voltages) in non-volatile memory cells rather than as digital values. This parameter change enables analog computation with lower energy consumption while the introduced output block ensures that the analog values can be accurately converted back to digital form when needed for verification or further processing.

Inventive Principle:

Principle #35Parameter changes

3Manufacturing precision

If non-volatile memory cells are programmed to hold specific charge values, then synaptic weight precision is improved, but the complexity of programming and verification mechanisms increases

Engineering Contradiction:

Improvesynaptic weight precisionVSAvoidprogramming and verification mechanism complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent merges the programming and verification functions into a unified output block that handles both operations. The same current-to-voltage conversion and analog-to-digital conversion mechanisms are used for both programming memory cells with precise weight values and verifying those values, reducing overall system complexity despite the high precision requirements.

Inventive Principle:

Principle #5Merging (Combining)

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 allows for precise and efficient programming and reading of memory cell values, improving the accuracy and energy efficiency of neural networks by enabling continuous and fine-tuned adjustments to synapse weights, thereby enhancing the performance of artificial neural networks.

Implementation Method 1

Each of the plurality of memory cells is configured to store a weight value corresponding to a number of electrons on the floating gate

Methodology Applied

Scientific EffectElectron storage in floating gate: Electrical Accumulator

Implementation Method 2

a current-to-voltage converter for receiving a sequence of currents from one or more selected non-volatile memory cells

Methodology Applied

Scientific EffectCurrent-to-voltage conversion: Ohm's Law

Implementation Method 3

an analog-to-digital converter for converting the voltage or the sequence of voltages into a plurality of output bits

Methodology Applied

Scientific EffectAnalog-to-digital conversion:

Data Source

PatentUS20220215239A1Digital output mechanisms for analog neural memory in a deep learning artificial neural network

Publication Date: 2022.07.07 SILICON STORAGE TECHNOLOGY INC

AI summary

Numerous embodiments for reading or verifying a value stored in a selected non-volatile memory cell in a vector-by-matrix multiplication (VMM) array in an artificial neural network are disclosed. The embodiments comprise various designs of input blocks for applying inputs to the VMM array during a read or verify operation and various designs of output blocks for receiving outputs from the VMM array during the read or verify operation.