Power Optimization Scheduler for Neural Network Voltage Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing power optimization technologies face challenges in efficiently managing power consumption for neural network (NN) models due to variations in voltage requirements across different operations and devices, leading to suboptimal power distribution and potential device degradation.
Innovation Solution
A power optimization scheduler that determines optimal voltage values for processing devices based on NN model information, adjusting voltage and frequency according to computation amounts and performing tasks, and updates these values based on usage patterns to minimize degradation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If a fixed voltage value is used for all operations, then device simplicity is maintained, but power consumption cannot be optimized for different neural network operators
Solution Approach 1:
The patent implements dynamic voltage adjustment by determining different voltage values for different neural network operators (e.g., SGEMM, adder tree) based on their computational characteristics. The voltage value is dynamically selected from multiple candidate values rather than using a fixed voltage, allowing the system to adapt power supply to match actual operational requirements and minimize energy consumption.
Solution Approach 2:
The system changes the voltage parameter based on the type of neural network operator being executed. By establishing correspondence between operator types and optimal voltage values, the system adjusts the voltage parameter dynamically to match the computational workload, thereby optimizing power consumption while maintaining operational efficiency.
2Use of energy by moving object
If voltage is adjusted for each operator type, then power consumption is optimized, but the complexity of voltage management increases
Solution Approach 1:
The patent segments the neural network operators into different categories (e.g., matrix multiplication operators, addition operators) and assigns different voltage values to each segment. This segmentation approach allows fine-grained power optimization by treating different operational types differently, while the segmentation itself provides a structured method for managing the complexity through categorization.
Solution Approach 2:
The system introduces an intermediary mechanism (voltage determination module) that mediates between the neural network operator execution and the power supply. This intermediary component handles the complexity of voltage selection and management, shielding the rest of the system from the intricacies of power optimization while still achieving energy efficiency.
3Use of energy by moving object
If process variations are not accounted for, then device manufacturing is simpler, but power consumption cannot be optimized for individual devices
Solution Approach 1:
The patent applies local quality optimization by determining voltage values specific to each device based on its process variations. Instead of using a universal voltage for all devices, the system tailors the voltage to match the specific characteristics of each device's neural network operator performance, thereby optimizing power consumption for individual devices while accounting for manufacturing variations.
Data Source
AI summary
An operating method of a power optimization scheduler is provided, where the operating method of a power optimization scheduler includes obtaining information regarding a neural network (NN) model, determining a voltage value for a task to be performed by at least one processing device, based on the obtained information regarding the NN model, and controlling a power management device to apply a voltage corresponding to the determined voltage value to the at least one processing device.