Serializing neural network computing unit

By using a neural network computing unit with serialized input and low-precision storage elements, the problem of limited efficiency in high-precision converter circuits in existing technologies is solved, and a more efficient computing process is achieved.

CN114902241BActive Publication Date: 2026-02-27ANAFLASH INC
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202080082888.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-12-01
Filing Date
2020-12-02
Publication Date
2026-02-27
Estimated Expiration
2040-12-02

AI Technical Summary

Technical Problem

Existing neural network computing units require complex high-precision digital-to-analog converter circuits and analog-to-digital converter circuits in their in-memory computing architecture, which limits their efficiency.

Method used

By employing a serialized input method, utilizing low-precision storage elements and a 1-bit sensing amplifier, the input is sequentially applied to the storage elements through a control circuit to generate an output current, which is then converted into binary code by an accumulator circuit, simplifying the signal conversion process.

Benefits of technology

This reduces reliance on high-precision converter circuits, improves computational efficiency and performance, and lowers the demand for high-precision synaptic components.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114902241B_ABST
    Figure CN114902241B_ABST
Patent Text Reader

Abstract

A serializing neural network computing unit. The computing unit includes a bit line; a memory array having a plurality of memory blocks, each memory block having one or more storage elements, each storage element connected to the bit line; a control circuit configured to sequentially apply a serializing input to the storage elements such that the storage elements sequentially generate outputs in response to the serializing input, wherein each output corresponds to a product of the input and a weight value stored in the storage element; and a set of reference current levels for the control circuit, each of the set of reference current levels having a particular amount of current to control the storage elements to generate a respective output current corresponding to the set of reference current levels.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Nonvolatile memory array logic

    US20150356006A1

  • Compute in memory circuits with time-to-digital computation

    US20190042160A1

  • Neural network circuits having non-volatile synapse arrays

    US20190164044A1

  • System and methods for mixed-signal computing

    WO2019182730A1