Heterogeneous Neural Network Computing for Energy-Aware Frequency Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network computation systems face challenges in reducing energy consumption while maintaining target execution times due to the use of heterogeneous computing devices, as existing frequency control methods often fail to consider the specific power consumption and execution time requirements of these devices.
Innovation Solution
A neural network computing system that dynamically adjusts the operating frequencies of heterogeneous computing devices based on feedback control mechanisms, using a frequency level combination table to minimize energy consumption while meeting performance levels, by analyzing execution time ratios and power consumption characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If heterogeneous computing devices are used to perform neural network computation, then computation capability and versatility are improved, but power consumption increases
Solution Approach 1:
The patent implements dynamic frequency adjustment for heterogeneous computing devices based on real-time workload analysis. The system continuously monitors execution progress and adapts operating frequencies of CPU, GPU, and NPU during runtime, transitioning from static to dynamic resource management. This resolves the contradiction by enabling high computation capability when needed while reducing power consumption during low-utilization periods.
Solution Approach 2:
The system changes operational parameters (frequency levels) of heterogeneous computing devices to optimize the balance between computation capability and power consumption. By adjusting frequency parameters dynamically based on workload characteristics and execution stage, the system achieves versatile computation performance while managing energy consumption efficiently across different operational states.
2Productivity
If operating frequency is increased to reduce execution time, then productivity is improved, but energy consumption increases
Solution Approach 1:
The patent applies partial frequency scaling by adjusting only the frequency levels of specific computing devices that are currently bottlenecking execution, rather than uniformly increasing all frequencies. The system identifies which devices need frequency boosts based on workload distribution and execution progress, applying frequency increases only where necessary to maintain productivity while minimizing energy consumption.
Solution Approach 2:
The system dynamically adjusts frequencies based on real-time execution monitoring, transitioning from static high-frequency operation to adaptive frequency management. By continuously assessing execution progress and workload characteristics, the system optimizes the trade-off between execution speed and energy consumption, achieving high productivity only when and where needed.
3Use of energy by moving object
If feedback control is implemented to optimize frequency levels, then energy efficiency is improved, but system complexity increases
Solution Approach 1:
The patent implements feedback control mechanisms that monitor execution progress, workload characteristics, and power consumption metrics to dynamically adjust frequency levels. The system uses feedback from performance monitoring to continuously optimize the operational state of heterogeneous computing devices, achieving improved energy efficiency through closed-loop control while managing complexity through modular implementation.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A neural network computing system includes a processor including heterogeneous computing devices, a memory, a memory controller that controls the memory, and a system bus that communicates between the processor and the memory controller. The processor generates a frequency level combination table representing frequency level combinations of hardware devices that include the heterogeneous computing devices, the memory controller, and the system bus, each of the frequency level combinations maximizing a decreasing amount of execution time relative to an increasing amount of energy consumption when a performance level of the neural network model is increased by one level, determines, a target performance level, selects a frequency level combination among the frequency level combinations according to the target performance level, and controls, based on a selected frequency level combination, the hardware devices while the neural network model is executed.