Sparsity-Based Neural Network Mapping for SoC Thermal Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep convolutional neural networks (DCNs) face inefficiencies and potential thermal runaway due to non-uniform temperature and power consumption across computing units in embedded IoT devices with limited resources, leading to performance throttling and shutdown risks.
Innovation Solution
Dynamic remapping of neural network partitions based on sparsity and thermal stress of computing units within a system-on-chip (SOC), allocating operations involving more sparse kernels to hotter units to reduce computational load and temperature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep convolutional neural networks are deployed on embedded IoT devices with limited resources, then image recognition and speech recognition capabilities are improved, but thermal runaway and performance throttling occur due to non-uniform temperature and power consumption across computing units
Solution Approach 1:
The system dynamically remaps neural network partitions to different computing units based on real-time thermal conditions and kernel sparsity characteristics. This dynamic adaptation allows the system to respond to changing thermal states, preventing thermal runaway while maintaining DCN deployment capability on resource-constrained embedded devices
Solution Approach 2:
The patent applies different processing strategies to different computing units based on their local thermal conditions. Computing units experiencing thermal stress receive different kernel assignments (prioritizing sparser kernels) compared to cooler units, creating localized quality variations that prevent system-wide thermal runaway while maintaining overall reliability
2Productivity
If computing units operate at high power to maintain processing speed, then productivity is improved, but thermal hotspots form causing performance throttling and shutdown risks
Solution Approach 1:
The system changes the parameter of kernel selection based on thermal conditions, dynamically switching between dense and sparse kernels. When thermal hotspots are detected, the system transitions to using sparser kernels on affected computing units, reducing power consumption and temperature while maintaining processing continuity
Solution Approach 2:
The neural network processing is segmented into multiple partitions that can be independently assigned to different computing units. This segmentation allows selective placement of computationally intensive partitions on cooler units while directing sparser partitions to thermally stressed units, balancing productivity and temperature management
3Ease of operation
If uniform kernel distribution is applied across computing units, then ease of operation is improved, but non-uniform temperature and power consumption cause thermal inefficiency
Solution Approach 1:
The kernel distribution strategy transitions from static uniform allocation to dynamic adaptive allocation based on thermal feedback. The system continuously monitors computing unit temperatures and adjusts kernel assignments accordingly, maintaining ease of operation through automated remapping while eliminating thermal inefficiency caused by uniform distribution
Solution Approach 2:
The system implements a feedback mechanism where thermal conditions of computing units are continuously monitored and used to inform subsequent kernel allocation decisions. This closed-loop control replaces simple uniform distribution with thermally-aware adaptive distribution, reducing energy loss while maintaining operational simplicity through automated decision-making
4Temperature
If sparse kernels are processed on thermally-stressed computing units, then temperature balancing is improved, but computational load must be dynamically adjusted
Solution Approach 1:
The system changes the sparsity parameter of selected kernels to match thermal conditions of computing units. By selecting sparser kernels for thermally-stressed units and denser kernels for cooler units, the system achieves temperature balancing while the automated parameter selection process manages the inherent complexity
Data Source
AI summary
A method for an artificial neural network includes receiving a set of input values to be convolved with a plurality of kernels via a plurality of computing units of a system-on-chip (SOC). A temperature associated with each of the plurality of computing units of the SOC is detected. The plurality of kernels are mapped to the plurality of computing units of the SOC based on the detected temperature associated with each of the plurality of computing units and a sparsity of each of the plurality of kernels. Convolution operations of the set of input values with the plurality of kernels are performed using the plurality of computing units. A kernel with a higher sparsity level is convolved with the set of input values on a computing unit associated with a greater temperature than a kernel with a lower sparsity level. An inference is generated based on the convolution operations.