Analog Neural Memory Decoders for Single-Cell Verify in VMM Arrays

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing decoding circuits for conventional flash memory arrays are not suitable for use with vector-by-matrix multiplication (VMM) arrays in analog neuromorphic memory systems, as they fail to accommodate the verify operation on a single selected memory cell while performing a read operation on all memory cells in the array.

Innovation Solution

The development of specialized decoders, including word line drivers and bit line decoders, that enable individual programming, erasing, and reading of memory cells in a VMM array without disturbing other cells, and support continuous analog programming.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional flash memory decoding circuits are used, then the system can perform read operations on all memory cells, but it cannot perform verify operations on a single selected memory cell without disturbing other cells

Engineering Contradiction:
Improveoperation mode flexibilityVSAvoidcell isolation accuracy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The decoding circuit is segmented into separate word line decoding paths: a first decoding path for selecting a single word line during programming/erasing operations, and a second decoding path for selecting multiple word lines simultaneously during read operations. This segmentation allows the circuit to adapt to different operation modes while maintaining proper cell isolation through the use of separate decoding mechanisms for each operation type.

Inventive Principle:
Principle #1Segmentation

2Reliability

If specialized decoders are developed for VMM arrays, then individual programming and reading can be performed without disturbing other cells, but the device complexity increases

Engineering Contradiction:
Improvecell isolation accuracyVSAvoiddecoder circuit complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The decoding circuit is designed with multi-functionality to handle both conventional flash memory operations and analog neuromorphic VMM operations using a unified architecture. The same decoding circuit can perform single-cell programming, array-wide reading, and analog weight tuning operations, reducing overall system complexity despite the specialized requirements of VMM arrays.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The decoding circuit employs dynamic control mechanisms that adjust its behavior based on the operation mode. Control signals dynamically switch between different decoding paths and selection schemes, allowing the circuit to adapt its complexity level to match the specific operation being performed, thereby optimizing performance without permanently increasing hardware complexity.

Inventive Principle:
Principle #15Dynamics

3Use of energy by moving object

If separate multiplication logic circuits are eliminated, then energy consumption is reduced, but the difficulty of detecting and measuring synapse weights increases

Engineering Contradiction:
Improveenergy consumptionVSAvoidsynapse weight measurement
Core Design Contradiction:
Use of energy by moving objectVSDifficulty of detecting and measuring

Solution Approach 1:

The patent introduces intermediary sensing circuits and reference cell structures that facilitate the measurement of analog synapse weights stored in the crossbar array. These intermediary components act as mediators between the analog memory cells and the digital readout system, enabling accurate weight detection without requiring separate multiplication logic circuits, thus maintaining energy efficiency while solving the measurement challenge.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4303772B1Decoders for analog neural memory in deep learning artificial neural network
Publication Date: 2025.11.12 SILICON STORAGE TECHNOLOGY INC
  • EP4303772B1 patent drawingFigure 1
  • EP4303772B1 patent drawingFigure 2
  • EP4303772B1 patent drawingFigure 3

AI summary

Numerous embodiments of decoders for use with a vector-by-matrix multiplication (VMM) array in an artificial neural network are disclosed. The decoders include bit line decoders, word line decoders, control gate decoders, source line decoders, and erase gate decoders. In certain embodiments, a high voltage version and a low voltage version of a decoder is used.