2D Router Mesh Segmentation for Neural Core Energy Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network architectures face challenges in achieving flexible and energy-efficient signal routing between neural cores, particularly in deep learning applications where arbitrary routing paths are required, leading to increased energy consumption due to larger capacitance in current routing methods.
Innovation Solution
The implementation of a flexible and energy-efficient 2-D router mesh with borderguard circuitry that includes three wire meshes: a plus-sign mesh, a horizontal-only mesh, and a vertical-only mesh, allowing for reduced energy usage by selectively routing signals across tiles and compute cores, and enabling arbitrary routing across a 2D data-mesh while supporting single-casting and multi-casting of data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional routing methods are used to achieve arbitrary routing paths between neural cores, then routing flexibility is improved, but energy consumption increases due to larger capacitance
Solution Approach 1:
The routing network is segmented into three distinct wire meshes: a plus-sign mesh for diagonal and orthogonal routing, a horizontal-only mesh for east-west routing, and a vertical-only mesh for north-south routing. This segmentation allows signals to be routed through optimized paths with lower capacitance, reducing energy consumption while maintaining routing flexibility.
Solution Approach 2:
The patent introduces a third dimension to the routing architecture by adding the horizontal-only and vertical-only meshes in addition to the plus-sign mesh. This multi-layer mesh structure provides multiple routing dimensions and paths, enabling arbitrary routing while selecting optimal low-capacitance paths to reduce energy consumption.
2Adaptability or versatility
If complex signal routing is implemented across multiple tiles, then routing capability is improved, but energy efficiency deteriorates
Solution Approach 1:
Different wire meshes are assigned to different routing needs: the plus-sign mesh handles diagonal and orthogonal routing, the horizontal-only mesh handles east-west routing, and the vertical-only mesh handles north-south routing. This local optimization ensures that each routing operation uses the most efficient path type, minimizing capacitance and energy loss even for complex multi-tile routing.
Solution Approach 2:
The routing system dynamically selects which wire mesh to use based on the specific routing requirements. Routers can choose between different mesh types and path configurations to optimize for energy efficiency while maintaining the ability to perform complex signal routing across multiple tiles when needed.
Data Source
AI summary
An array of neural cores has at least two dimensions. Each of the neural cores comprises ordered input wires, ordered output wires, and synapses, each of the synapses operatively coupled to one of the input wires and one of the output wires. Signal wires are provided. At least one of the signal wires is disposed along each dimension of the array of neural cores. Each of the signal wires is disposed along at least one dimension of the array. Routers are provided, each of which is operatively coupled to (i) one of the neural cores and (ii) at least two of the signal wires, one along each of the dimensions of the array of neural cores. Each of the routers is configured to selectively route a signal from one of its at least two coupled signal wires to its coupled neural core. Each of the routers is configured to selectively route a signal from its coupled neural core to one of its at least two coupled signal wires.