Access Official Technical PDF

Share Summary:LinkedInEmail

Neuromorphic computing is an approach to AI hardware that models aspects of how biological neurons communicate. For industrial automation, its potential significance lies in processing vision and sensor inputs at the edge, where equipment may need to interpret events without continuously sending all data to remote computing resources. The concept differs from conventional neural-network processing in how information is represented and when computation occurs.

BrainChip’s Akida technology, described in a 2019 company coverage report, provides an example of a neuromorphic System-on-a-Chip (NSoC) based on spiking neural networks (SNNs). The report presented industrial applications such as vision-guided robotics and the Industrial Internet of Things (IIoT) as potential target areas. These were prospective applications, not evidence of deployment or validated performance in a particular plant.

How neuromorphic computing processes information

Many familiar deep-learning systems use neural networks such as convolutional neural networks (CNNs). These process data through layers of calculations. A spiking neural network instead represents input as discrete events, or spikes. In a vision application, for example, a system may receive information derived from changes in an image rather than treating every input as an uninterrupted stream of equivalent data.

In the Akida architecture described in the report, artificial neurons are connected through artificial synapses. A synapse applies a threshold: signals that meet relevant conditions can propagate through the network, while others may not. The report also describes learning through changes to synaptic weights, a process associated with spike-timing-dependent plasticity. In practical terms, connections adjust based on patterns in the incoming spikes.

This event-driven design aims to make processing dependent on the presence of events. By contrast, a conventional network may continue processing successive input data whether or not meaningful changes are present. The distinction matters for applications where the input is continuous but useful decisions depend on detecting specific objects, movements or patterns.

Why edge processing matters in industrial automation

Factories generate data from cameras, sensors, robots and other connected equipment. Some applications require analysis near the source—for example, a vision system associated with a robot or an inspection station. Edge processing places computation on or near the device collecting data, rather than relying entirely on a remote server or cloud connection.

Neuromorphic hardware is intended to combine neural-network processing with a chip architecture designed around events. The report attributes potential speed and power advantages to parallel spike processing and the storage of learned information in synaptic weights. It also describes reduced reliance on external memory transfers as a possible contributor to lower energy use. These are architectural propositions; actual results depend on the implementation, workload and comparison conditions.

For automation teams, the possible areas of fit include:

  • Industrial machine vision: identifying objects or visual patterns in applications such as inspection and equipment monitoring.
  • Vision-guided robotics: interpreting visual input relevant to a robot’s task.
  • IIoT edge devices: analyzing selected sensor or image inputs where local processing is useful.
  • Mobile or power-constrained equipment: exploring AI functions in devices that cannot depend on substantial computing infrastructure.

These examples describe application categories, not guaranteed capabilities. Any implementation still needs to meet the specific requirements for accuracy, timing, integration and operating conditions.

From development environment to production integration

A new AI chip is only one part of an automation solution. Models must be prepared, tested, integrated with the surrounding system and evaluated against the intended task. The report describes the Akida Development Environment as a software simulation for creating, training and testing networks, and for running inference before silicon samples were expected to be available at the time.

That development sequence illustrates a broader integration consideration: engineering teams may need to assess a platform before hardware is ready for production design. Once a device is available, a design-in process can involve adapting the chip or its intellectual property to a customer’s system, followed by testing before commercial deployment. The report also outlines training before deployment, with learned synaptic weights transferred to other chips. It notes that a deployed chip might continue fine-tuning weights, but that behavior would need to be assessed for the application and operational controls in question.

For plant and controls engineers, the evaluation should begin with the use case rather than the architecture. Define what the system must recognize, what data it will receive, where inference will run and how results will connect to the existing automation system. Then determine whether the development tools, deployment workflow and available hardware support those requirements.

Questions for technical decision-makers

Before considering neuromorphic hardware for an industrial project, teams can use a structured assessment:

  1. Specify the task: identify the objects, events or patterns the system must detect.
  2. Set acceptance criteria: establish how accuracy and response behavior will be assessed for the actual application.
  3. Check integration needs: examine interfaces with cameras, sensors, controllers and other system components.
  4. Evaluate training and updates: clarify how a model is prepared, transferred, tested and maintained.
  5. Compare the complete solution: consider the hardware, software, development effort and operating requirements, not just chip-level claims.

The source report also distinguishes between supplying complete NSoC devices and licensing selected intellectual-property blocks. That distinction is relevant to system planning: an automation OEM or device manufacturer may evaluate a complete component, while a semiconductor partner may consider incorporating IP into another design. The commercial route does not, by itself, determine whether a technology is suitable for a control or vision task.

Conclusion: assess the workload, not the label

Neuromorphic computing offers a different way to organize neural-network processing, using spikes and event-driven activity rather than treating every input as a continuous stream of conventional calculations. For industrial automation, its potential relevance is clearest in edge applications involving vision, robotics and IIoT devices, where teams are evaluating local AI processing.

The practical decision is whether an implementation can meet the requirements of a specific operation and integrate reliably with its equipment and software environment. Engineers should validate the target workload, development process and system interfaces before treating architectural promises as production benefits. Neuromorphic computing is a technology option to investigate—not a substitute for application-level testing.

Access Official Technical PDF

About The Author