In-Memory Computing Device Retains Reused Elements
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increased accuracy of artificial neural networks leads to higher computation, memory access, and data movement requirements, resulting in reduced system performance, power consumption, and speed due to the need for extensive data processing in computing systems.
Innovation Solution
A data processing system with an in-memory computing device that includes multiple processing elements, where a scheduler minimizes data movement by retaining reused elements within processing elements and distributing new elements efficiently, allowing neural network operations to be performed without moving reused elements between processing elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If accuracy of artificial neural network is increased, then problem solving ability is improved, but computation amount and data movement are increased causing reduction in speed and power consumption
Solution Approach 1:
The input feature map is divided into multiple division maps, and the weight filter is divided into multiple weight map groups. This segmentation allows parallel processing across multiple processing elements while reducing the data movement burden on each individual element, thereby maintaining high accuracy neural network operations at improved processing speeds
Solution Approach 2:
A scheduler is introduced as an intermediary component that manages the distribution of division maps to processing elements. The scheduler identifies reused elements and ensures they are retained in the same processing element, acting as a mediator between the data flow requirements and the processing element architecture to minimize unnecessary data movement
2Measurement precision
If accuracy of artificial neural network is increased, then problem solving ability is improved, but power consumption is increased due to extensive data processing
Solution Approach 1:
The scheduler performs preliminary identification of reused elements before the neural network operation begins. By pre-determining which elements will be reused and assigning them to specific processing elements in advance, the system avoids unnecessary data movement during execution, thereby reducing power consumption while maintaining accurate computation
Solution Approach 2:
The scheduler acts as an intermediary that optimizes the interaction between data requirements and processing elements. It minimizes power consumption by intelligently managing data distribution and retention strategies, ensuring that reused elements remain in processing elements rather than being repeatedly transferred from memory
3Measurement precision
If accuracy of artificial neural network is increased, then problem solving ability is improved, but system performance is deteriorated due to reduced speed and increased power consumption
Solution Approach 1:
By segmenting the neural network operation into multiple division maps processed by multiple processing elements in parallel, the system achieves both high accuracy and improved performance. The segmentation enables efficient resource utilization while maintaining computational precision
Solution Approach 2:
The scheduler serves as a performance-optimizing intermediary that coordinates the parallel processing elements. It ensures that reused elements are properly managed across the segmented processing architecture, thereby maintaining system reliability and performance while supporting high-accuracy neural network operations
Data Source
AI summary
A data processing system includes a controller configured to receive a neural network operation processing request from a host device; and an in-memory computing device including a plurality of processing elements. The in-memory computing device is configured to receive an input feature map and a weight filter from the controller, and perform a neural network operation in the plurality of processing elements based on the weight filter and a plurality of division maps generated from the input feature map, wherein the in-memory computing device performs the neural network operation by not moving a reused element, which is operated at least twice among elements constituting the division maps during the neural network operation, between the processing elements.