DNN Workload Partitioning for Lower-Power Parallel Neuron Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) require high computational complexity, making general-purpose processors like GPUs unsuitable for long-duration tasks in low-power devices such as smartphones and power-over-Ethernet security cameras, where reduced power consumption is critical.
Innovation Solution
A DNN module that dynamically partitions neuron workload and assigns these partitions to groups of neurons for parallel processing, allowing for faster completion and quicker power-down, thereby reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If general-purpose processors like GPUs are used to provide high computational performance for DNN operations, then processing speed and accuracy are improved, but power consumption increases making them unsuitable for battery-powered portable devices
Solution Approach 1:
The patent segments the DNN processing workload into multiple partitions and assigns them to different groups of neurons for parallel processing. This segmentation allows the system to complete processing faster and power down more quickly, reducing overall power consumption while maintaining high computational performance through parallel execution of partitioned workloads
Solution Approach 2:
The patent implements dynamic workload partitioning where the system adaptively divides and assigns DNN processing tasks to neuron groups based on available resources and processing requirements. This dynamic approach optimizes the balance between computational performance and power consumption by adjusting the degree of parallelization and processing depth according to real-time conditions
2Productivity
If DNN processing is performed continuously in battery-powered devices, then useful computational tasks are completed, but battery life is reduced due to high power consumption
Solution Approach 1:
The patent employs periodic action by processing DNN workloads in discrete, partitioned batches rather than continuous processing. Each batch is processed by assigned neuron groups and then the system can enter a low-power state, creating periodic processing cycles that reduce average power consumption while maintaining necessary DNN processing capability over extended periods
Solution Approach 2:
By segmenting the continuous DNN processing into discrete partitioned tasks that can be completed and powered down between executions, the system extends battery life while preserving DNN processing capability for when power is available
3Speed
If more neurons are used to process DNN workloads in parallel, then processing speed increases, but the complexity of workload management and scheduling increases
Solution Approach 1:
The patent manages the complexity of coordinating multiple neurons by segmenting the workload into clearly defined partitions with explicit assignment to specific neuron groups. Each partition represents an independent, manageable unit of work with well-defined data dependencies, making the scheduling and coordination of multiple neurons tractable while still achieving high parallel processing speeds
Data Source
AI summary
A deep neural network (DNN) module is disclosed that can dynamically partition neuron workload to reduce power consumption. The DNN module includes neurons and a group partitioner and scheduler unit. The group partitioner and scheduler unit divides a workload for the neurons into partitions in order to maximize the number of neurons that can simultaneously process the workload. The group partitioner and scheduler unit then assigns a group of neurons to each of the partitions. The groups of neurons in the DNN module process the workload in their assigned partition to generate a partial output value. The neurons in each group can then sum their partial output values to generate a final output value for the workload. The neurons can be powered down once the groups of neurons have completed processing their assigned workload to reduce power consumption.