Processor and control method thereof

CN111133455BActive Publication Date: 2024-12-31SAMSUNG ELECTRONICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201880061613.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2017-10-31
Filing Date
2018-10-04
Publication Date
2024-12-31
Estimated Expiration
2038-10-04

AI Technical Summary

Technical Problem

When existing processors perform deep learning tasks, it is difficult to effectively reduce the amount of peak data movement from external memory, resulting in limited scalability of the processor, especially when multiple processors implement deep neural networks (DNNs). Path activity is very bursty and difficult to achieve through simple parallel connections.

Method used

A processor is designed, which includes a plurality of processing elements arranged in matrix form, first and second data input units, and a shifter through which instructions and data are input and processed between multiple rows of processing elements , reduce the amount of data movement between the processor and external memory, and improve the efficiency of the data path.

Benefits of technology

With this design, the amount of peak data movement from external memory to the processor is significantly reduced, the processor is improved scalability, the hardware design is simplified, and the processor integration and scalability is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111133455B_ABST
    Figure CN111133455B_ABST
Patent Text Reader

Abstract

A processor for performing deep learning is provided herein. The processor includes a processing element unit, which includes a plurality of processing elements arranged in a matrix form, the matrix form including a first row of a plurality of processing elements and a second row of a plurality of processing elements. The processing element is fed with filter data by a first data input unit connected to the first row of the plurality of processing elements. The second data input unit feeds target data to the processing element. A shifter composed of registers feeds instructions to the processing element. A controller in the processor controls the processing element, the first data input unit and the second data input unit to process the filter data and the target data, thereby providing a product-sum (convolution) function.
Need to check novelty before this filing date? Find Prior Art