Adaptive NoC Construction via ML Traffic Flow Sorting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Network on Chip (NoC) designs face challenges in optimizing channel bandwidth and latency due to scalability limitations, traffic profile uncertainties, and hardware reconfigurability, leading to sub-optimal power consumption and performance.
Innovation Solution
The method involves using machine learning algorithms to determine traffic flow groups, sorting orders, and mapping algorithms to generate an efficient NoC architecture that can adapt to varying traffic profiles by allocating routes, virtual channels, and layers, optimizing channel capacity and host placement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional bus or crossbar interconnects are used, then implementation is simple, but scalability is limited
Solution Approach 1:
The interconnect is segmented into multiple independent Network-on-Chip (NoC) layers, each handling specific traffic flows. This segmentation enables scalable architecture by allowing each layer to be independently designed and configured, while collectively providing the required bandwidth and routing capabilities for large-scale systems.
Solution Approach 2:
The patent introduces a vertical dimension by stacking multiple NoC layers, transforming the traditional two-dimensional interconnect into a three-dimensional architecture. This dimensional extension allows simultaneous handling of multiple traffic patterns across different layers, dramatically improving scalability without proportionally increasing routing complexity.
2Reliability
If deterministic routing is used, then packet ordering is maintained and deadlocks are avoided, but load balancing across path diversities is not achieved
Solution Approach 1:
The system dynamically selects routing strategies based on traffic flow characteristics. For traffic patterns requiring strict ordering, deterministic routing is applied within individual layers. For traffic patterns benefiting from load balancing, the system utilizes multiple layers with different routing algorithms, achieving both reliability and productivity through adaptive configuration.
Solution Approach 2:
Different routing strategies are segmented and applied to different traffic flows across multiple layers. This allows simultaneous execution of deterministic routing for order-critical traffic and adaptive load-balanced routing for performance-critical traffic, resolving the contradiction between reliability and productivity.
3Loss of time
If shortest path routing is used, then latency is minimized, but adaptability to traffic profile uncertainties is reduced
Solution Approach 1:
The system dynamically adapts routing paths based on actual traffic profiles and network conditions. While shortest path routing provides low latency for predictable traffic, the multi-layer architecture enables real-time reconfiguration to alternative paths when traffic patterns change or congestion occurs, maintaining both low latency and high adaptability.
Solution Approach 2:
The patent changes routing parameters dynamically based on traffic flow characteristics. Different layers can employ different routing metrics (e.g., hop count, bandwidth availability, latency) to optimize performance for specific traffic types, achieving both minimal latency and adaptability to uncertain traffic profiles.
4Productivity
If NoC is designed for specific traffic profiles, then performance is optimized, but hardware reconfigurability is reduced
Solution Approach 1:
The multi-layer NoC architecture provides universal functionality by designing each layer to handle multiple traffic patterns. Rather than dedicating specific hardware to single traffic profiles, each layer can be configured to support various routing algorithms and traffic types, achieving both optimized performance and hardware reconfigurability through multi-functional design.
Solution Approach 2:
The system dynamically reconfigures layer assignments and routing parameters based on actual traffic demands. This allows the hardware to maintain optimized performance for current traffic profiles while retaining the capability to reconfigure for different traffic patterns, resolving the contradiction between performance optimization and reconfigurability.
5Quantity of substance
If more layers are added to increase bandwidth, then channel capacity increases, but device complexity increases
Solution Approach 1:
The patent segments traffic flows across multiple layers, with each layer handling a subset of traffic patterns. This segmentation increases total channel bandwidth by utilizing vertical space, while managing complexity through systematic organization of layers and automated traffic assignment algorithms that prevent exponential growth in configuration complexity.
Solution Approach 2:
Rather than providing full bandwidth capacity in a single complex layer, the system uses multiple layers with partial bandwidth each. This approach achieves total required bandwidth while keeping individual layer complexity manageable, as each layer can be independently optimized and configured.
Data Source
AI summary
Example implementations described herein are directed to systems and methods for generating a Network on Chip (NoC), which can involve determining a plurality of traffic flows from a NoC specification; grouping the plurality of traffic flows into a plurality of groups; utilizing a first machine learning algorithm to determine a sorting order on each of the plurality of groups of traffic flows; generating a list of traffic flows for NoC construction from the plurality of groups of traffic flows based on the sorting order; utilizing a second machine learning algorithm to select one or more mapping algorithms for each group of the plurality of groups of traffic flows for NoC construction; and generating the NoC based on a mapping from the selection of the one or more mapping algorithms.