Neural Network Workload Re-allocation Across Heterogeneous Processors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In processing environments like automotive systems, existing thermal mitigation techniques for SoCs often compromise performance by reducing processor voltage and frequency, which is undesirable in safety-critical scenarios, and fail to dynamically allocate workloads to minimize thermal effects effectively.
Innovation Solution
Implementing a workload re-allocation system that dynamically migrates neural network units across heterogeneous processors in an SoC based on temperature and performance metadata, identifying optimal target processors to balance thermal dissipation and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Temperature
If thermal mitigation techniques reduce processor voltage and frequency, then temperature increases are controlled, but performance decreases
Solution Approach 1:
The patent implements dynamic workload allocation that adjusts processor assignments in real-time based on thermal conditions and performance requirements. The system continuously monitors temperature and dynamically migrates neural network workloads between processors, enabling adaptive response to thermal conditions without static performance degradation. This dynamic approach allows the system to maintain high performance on cool processors while managing thermal loads, resolving the contradiction between temperature control and performance maintenance.
Solution Approach 2:
The patent applies local quality by distributing workloads across heterogeneous processors with different thermal characteristics and performance capabilities. Instead of uniformly throttling all processors, the system identifies specific processors with acceptable thermal conditions and assigns workloads to those local regions. This selective allocation allows hot processors to be cooled while maintaining high performance on suitable processors, addressing both thermal mitigation and performance requirements simultaneously.
2Productivity
If neural network software is configured to run on a specific processor type, then execution efficiency is optimized, but adaptability to thermal conditions is reduced
Solution Approach 1:
The patent transforms static processor configuration into dynamic workload allocation. Instead of fixing neural network software to specific processor types at compile time, the system continuously evaluates thermal conditions and performance metrics at runtime, then dynamically assigns workloads to appropriate processors. This dynamic mechanism maintains execution efficiency by selecting suitable processors while providing adaptability to changing thermal conditions, resolving the contradiction between optimization and flexibility.
Solution Approach 2:
The patent implements a universal workload allocation system that can assign neural network workloads to any suitable processor type (CPU, GPU, NPU, DSP) based on real-time conditions. The heterogeneous processor pool serves multiple functions, and the allocation system universally manages workload distribution across all processor types. This multi-functionality allows the system to maintain optimized execution by selecting the most appropriate processor for each workload while adapting to thermal conditions, overcoming the limitation of fixed processor assignments.
Data Source
AI summary
Neural network workload re-allocation in a system-on-chip having multiple heterogenous processors executing one or more neural network units may be based on measurements associated with the processors' conditions and on metadata associated with the neural network units. Metadata may be contained in an input file along with neural network information. Measurements characterizing operation of the processors may be obtained and compared with one or more thresholds. A neural network unit executing on a processor may be identified as a candidate for re-allocation based on metadata associated with the neural network unit and results of the comparisons. A target processor may be identified based on the metadata and results of the comparisons, and the candidate neural network neural network unit may be re-allocated to the target processor.