Dynamic Neural Network Accelerator Configuring Operation Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks face constraints in development speed and energy consumption due to the large number of operations, with conventional accelerator architectures lacking flexibility and failing to improve operation speed or reduce operation quantity.
Innovation Solution
An operation device and method that dynamically configures operation units based on data characteristics, allowing for shutdown of multipliers and adders when data meets certain conditions, reducing power consumption without affecting operation speed, and utilizing data processing to compress or expand data accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional accelerator architectures perform targeted acceleration designs by analyzing the most time-consuming operation part, then operation speed is improved, but device complexity increases and flexibility is reduced
Solution Approach 1:
The accelerator architecture is segmented into multiple operation groups, each handling specific operation types. This allows the system to divide complex neural network operations into manageable segments that can be processed independently, improving overall operation speed while maintaining modular complexity management.
Solution Approach 2:
The accelerator employs dynamic configuration where operation groups and their internal components (multipliers, adders) can be dynamically enabled or disabled based on the specific computation task. This dynamic adaptability improves operation speed for targeted operations while reducing effective device complexity by deactivating unnecessary components.
2Productivity
If the quantity of operations in deep neural networks is increased, then computational capability is improved, but energy consumption increases
Solution Approach 1:
The accelerator applies local quality by configuring different numbers of multipliers and adders in different operation groups based on the specific computational requirements of each neural network layer. This allows high computational capability where needed while reducing energy consumption in areas where fewer operations are required.
Solution Approach 2:
The system changes operational parameters by dynamically adjusting the number of active operation units (multipliers and adders) based on the sparsity and complexity of the input data. When data sparsity is detected, the system reduces the number of active operations, thereby maintaining computational capability for essential operations while reducing overall energy consumption.
3Productivity
If accelerator components are continuously activated to maintain operation speed, then processing throughput is improved, but power consumption increases
Solution Approach 1:
The accelerator implements dynamic power management by continuously monitoring data characteristics and adjusting the activation state of operation groups accordingly. When sparse or redundant operations are detected, corresponding multipliers and adders are deactivated, reducing power consumption while maintaining processing throughput for essential computations through dynamic reallocation of active components.
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
A computing device, comprising: a computing module, comprising one or more computing units; and a control module, comprising a computing control unit, and used for controlling shutdown of the computing unit of the computing module according to a determining condition. Also provided is a computing method. The computing device and method have the advantages of low power consumption and high flexibility, and can be combined with the upgrading mode of software, thereby further increasing the computing speed, reducing the computing amount, and reducing the computing power consumption of an accelerator.