Memory Device with Selection Module for Neural Network MAC Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory devices struggle to efficiently perform data operations for deep neural networks, particularly in handling matrix and vector data for operations like multiply and accumulate (MAC), due to limitations in processing in memory (PIM) and utilization of MAC units.
Innovation Solution
A memory device with an operation module that includes multiple operators for performing data operations, a register for storing vector data, and a selection module that maps data to operators based on various operation modes, allowing for efficient handling of matrix and vector data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If processing in memory (PIM) is used to perform MAC operations, then computational efficiency is improved, but the device complexity increases due to integration of operators and selection modules
Solution Approach 1:
The patent merges memory storage functions with data processing functions by integrating operators directly into the memory device. The operation module combines multiple operators and a selection module within the memory architecture, enabling MAC operations to be performed directly on stored data without external processing, thus improving computational efficiency while managing complexity through functional integration
Solution Approach 2:
The memory device is designed with multi-functionality by incorporating operators that can perform various data operations including MAC operations. The selection module enables the same hardware structure to adapt to different operation modes and data formats, allowing the device to serve both as memory storage and as a processing unit for neural network computations
2Adaptability or versatility
If multiple operation modes are supported for different data dimensions, then adaptability is improved, but the control complexity increases
Solution Approach 1:
The selection module implements dynamic adaptability by receiving operation mode inputs that determine how data is mapped to operators. The system can dynamically adjust its operation based on input dimensions, supporting different modes such as full data mapping, partial data selection, or specific data element processing, allowing the same hardware to efficiently handle varying neural network operation requirements without physical reconfiguration
Data Source
AI summary
A memory device and a method of operating the same are provided. The memory device includes an operation module including a plurality of operators configured to perform a data operation, a register configured to store a plurality of pieces of data, and a selection module configured to receive an operation mode input and map the plurality of pieces of data stored in the register to the plurality of operators based on the operation mode input.