Method for determining quantization parameters in neural networks and related products

JP7578493B2Active Publication Date: 2024-11-06SHANGHAI CAMBRICON INFORMATION TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2020573207
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2019-06-27
Filing Date
2019-09-19
Publication Date
2024-11-06
Estimated Expiration
2039-09-19

AI Technical Summary

Technical Problem

Neural networks require high-precision data formats like floating-point numbers, which occupy significant storage space and increase processing speed and bandwidth requirements, leading to higher costs and resource consumption.

Method used

A method and device for adjusting data bit width through quantization, converting high-precision data to low-precision fixed-point numbers, allowing neural network operations to reduce storage space and improve computing performance by determining optimal bit widths based on quantization error and statistical data analysis.

Benefits of technology

This approach reduces the space required for neural network storage and improves computing efficiency by converting data to fixed-point numbers, maintaining accuracy while minimizing resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007578493000043
    Figure 0007578493000043
  • Figure 0007578493000044
    Figure 0007578493000044
  • Figure 0007578493000045
    Figure 0007578493000045
Patent Text Reader

Abstract

This technical solution relates to a board card, which includes a storage device, an interface device, a control device, and an AI chip, where the AI ​​chip and the storage device, the control device, and the interface device are connected respectively. The storage device is for storing data. The interface device is for implementing data transmission between the AI ​​chip and external devices. The control device is for monitoring the status of the AI ​​chip. The board card may also be for performing AI calculations. [Selection diagram] Figure 12
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Encoding system and decoding system

    JP1998233691A

  • Adaptive bit-width reduction for neural networks

    US20190050710A1