Neural Network Adder Multiplexer Controller for Reduction Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network operations face inefficiencies when processing large spatial dimension input feature maps, requiring extensive hardware and control systems for reduction operations, which can lead to increased overhead and complexity.
Innovation Solution
A neural network operation apparatus and method that includes an adder, multiplexers, and a controller to manage data processing, utilizing reset and phase signals to control operations and reduce overhead, enabling efficient reduction operations by separating processing logic and storing data based on channel indices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If hardware is added to perform reduction operation, then processing capability is improved, but device complexity increases
Solution Approach 1:
The main datapath is designed to perform both convolution operations and reduction operations using the same hardware resources. The adder tree structure can accumulate data for reduction operations while also performing convolution calculations, eliminating the need for separate dedicated reduction hardware and thus avoiding increased device complexity.
2Ease of operation
If control system is added to control main datapath and reduction module, then operational control is improved, but device complexity increases
Solution Approach 1:
The control logic for the reduction operation is merged into the existing control system that manages the main datapath. The controller generates control signals that coordinate both convolution and reduction operations through a unified control mechanism, avoiding the need for separate control systems and reducing overall device complexity.
Solution Approach 2:
The reduction operation is performed in periodic phases within the main datapath execution cycle. The controller activates reduction operations at specific intervals or phases, allowing the same control system to manage both continuous convolution operations and periodic reduction operations without requiring additional dedicated control hardware.
3Adaptability or versatility
If support of compiler or software is added, then operational flexibility is improved, but device complexity increases
Solution Approach 1:
The main datapath is designed with universal functionality to handle both convolution and reduction operations, allowing the compiler or software to generate appropriate control signals for different operation types without requiring hardware modifications. This software-controlled approach provides operational flexibility while avoiding the complexity of dedicated hardware for each operation type.
Data Source
AI summary
A neural network operation apparatus and method are disclosed. The neural network operation apparatus may include an adder configured to perform addition of data for performing a neural network operation and main data, a first multiplexer configured to output one of an output result of the adder and the main data based on a reset signal, a second multiplexer configured to output one of the main data and a quantization result of the data based on a phase signal, and a controller configured to control the first and second multiplexers by generating the reset signal and the phase signal.