Neuron Placement Optimization for Neuromorphic Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In neuromorphic systems, the cumulative delivery delay of spike packets across neurons is not optimized, leading to inefficiencies in processing time and energy consumption due to sequential neuron processing and varying spike travel times across a 2D grid of neurosynaptic cores.
Innovation Solution
The neurons are reordered based on their relative delivery delays, with those requiring the longest time to travel being placed earliest in the processing sequence to minimize cumulative delivery delay, and additional programmatic delays are considered to prioritize neurons with shorter delays, using a heuristic approach to optimize spike delivery within a millisecond window.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If neurons are processed in sequential order on a 2D grid of neurosynaptic cores, then the system maintains simplicity in neuron arrangement, but cumulative delivery delay increases due to varying spike travel times
Solution Approach 1:
The patent applies preliminary action by calculating and determining the optimal processing order of neurons before actual spike delivery occurs. The system pre-computes delivery delays for all neuron pairs and establishes an optimized processing sequence in advance, allowing neurons to be processed in an order that minimizes cumulative delay without requiring complex real-time adjustments during spike delivery
Solution Approach 2:
The patent changes the parameter of neuron processing order from a fixed sequential sequence to an optimized permutation based on calculated delivery delays. By varying the processing order parameter according to spike travel times between neurons, the system reduces cumulative delivery delay while maintaining the same physical 2D grid architecture
2Ease of operation
If neurons with longer spike travel times are processed later in the sequence, then processing follows a simple sequential pattern, but energy consumption increases due to extended processing time
Solution Approach 1:
The system performs preliminary calculation of spike delivery delays for all neuron pairs before processing begins. This pre-computation allows the system to establish an optimized processing order that minimizes total energy consumption by ensuring neurons requiring longer delivery times are processed earlier, avoiding wasted energy from extended processing sequences
3Device complexity
If neurons are arranged without optimization, then the neural network description remains simple, but spike arrival reliability decreases due to delays exceeding the millisecond window
Solution Approach 1:
The patent applies preliminary action by computing delivery delays and determining optimal neuron processing order before spike delivery occurs. This pre-planning ensures that neurons are processed in an order that guarantees spike arrivals within the required millisecond window, improving reliability without adding complexity to the neural network description itself
Solution Approach 2:
The system uses the neural network description itself to compute delivery delays and determine optimal processing order, without requiring external control mechanisms. The optimization is self-contained within the existing system framework, improving spike arrival reliability while maintaining description simplicity
Data Source
AI summary
Neuron placement in a neuromorphic system to minimize cumulative delivery delay is provided. In some embodiments, a neural network description describing a plurality of neurons is read. A relative delivery delay associated with each of the plurality of neurons is determined. An ordering of the plurality of neurons is determined to optimize cumulative delivery delay over the plurality of neurons. An optimized neural network description based on the ordering of the plurality of neurons is written.