Switching Network for Convolutional Neural Network Data Distribution

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing hardware implementations of spiking convolutional neural networks face challenges with parallel data access due to routing and concurrency issues, particularly when using multistage interconnection networks (MINs) that are not compatible with analogue signals and lack flexibility in handling circular shifts.

Innovation Solution

A device with a switching network controlled by a control unit, capable of applying circular shifts to convolution kernel coefficients, dynamically configures switches in response to input events and kernel parameters, allowing parallel data access in a convolutional neural network processor.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If multistage interconnection networks (MINs) are used for parallel data access, then routing capability is improved, but compatibility with analogue signals is lost and flexibility for circular shifts is reduced

Engineering Contradiction:

Improveparallel data access speedVSAvoidflexibility for circular shifts

Core Design Contradiction:

SpeedVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamic switching networks that can reconfigure their connection topology based on the required circular shift amount. Unlike static MINs, these switches can adapt their routing paths dynamically, allowing the same hardware to support variable circular shifts while maintaining parallel access capability. The switching network changes its internal configuration to route data through different paths corresponding to the desired shift amount.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The invention changes the operational parameters of the switching network to accommodate both parallel access and circular shift requirements. By modifying the routing parameters and switch configuration states based on the input event and kernel parameters, the system achieves flexibility without sacrificing speed. The switching network operates in different modes depending on the required operation.

Inventive Principle:

Principle #35Parameter changes

2Quantity of substance

If separate digital memory is used to store convolution kernel coefficients, then storage capability is improved, but memory access complexity and routing overhead increase

Engineering Contradiction:

Improvestorage capacity for kernel coefficientsVSAvoidmemory access and routing complexity

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent merges the memory storage function with the switching network by implementing distributed memory architecture where kernel coefficients are stored locally at or near the processing elements. This integration eliminates the need for complex centralized memory access and reduces routing overhead, as the coefficients are already positioned where they are needed for computation.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The system performs preliminary distribution of kernel coefficients to the appropriate processing elements before the convolution operation begins. This pre-positioning of data eliminates the need for complex real-time routing during computation, as the coefficients are already in place and can be accessed directly by the processing elements.

Inventive Principle:

Principle #10Preliminary action

3Ease of operation

If conventional switching networks are used for data distribution, then routing capability is improved, but concurrency of access to data creates problems

Engineering Contradiction:

Improvedata distribution capabilityVSAvoidconcurrency access reliability

Core Design Contradiction:

Ease of operationVSReliability

Solution Approach 1:

The patent segments the data distribution network into multiple independent switching stages and pathways. By dividing the routing function across multiple segments, concurrent access requests can be handled in parallel without interference. Each segment operates independently, allowing simultaneous data flows to proceed without conflicts or race conditions.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The switching network acts as an intermediary layer between memory and processing elements, managing concurrent access requests through coordinated switching. The intermediary switches buffer and route multiple data streams simultaneously, resolving potential conflicts before data reaches the processing elements, thereby ensuring reliable concurrent access.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS11423296B2Device and method for distributing convolutional data of a convolutional neural network

Publication Date: 2022.08.23 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES

AI summary

A device for distributing the convolution coefficients of the least one convolutional kernel of a convolutional neural network is provided, the coefficients being carried by an input bus, to a set of processing units in a processor based on a convolutional-neural-network architecture. The device comprises at least one switching network that is controlled by at least one control unit, the switching network comprising a set of switches that are arranged to apply circular shifts to at least one portion of the input bus. For each convolution kernel, each control unit is configured to dynamically control certain at least of the switches of the switching networks in response to an input event applied to the convolution kernel and at least one parameter representing the maximum size of the convolution kernels.