BrainChip’s 2019 year-end update set out a transition for the Akida neuromorphic processor: from design and verification toward fabrication, engineering samples and commercialization. For automation engineers and OEM teams, the announcement highlighted two routes to adoption—licensing Akida intellectual property for custom ASICs and evaluating the processor as a standalone device. It also described the event-based neural network approach behind the platform and its potential fit in edge applications involving vision and audio.
Akida neuromorphic processor development status
BrainChip reported that Akida’s logic and layout designs were complete. Design-for-Test work was still to be completed, with wafer fabrication at Taiwan Semiconductor Manufacturing Company (TSMC) planned to begin in January 2020. During the fabrication cycle, the company planned to complete test programs and hardware. Software and supporting materials were also in development, with availability expected alongside engineering samples.
The update described Akida as a neural network device supporting on-chip training, inference and incremental learning. The design was being simulated in the Akida Development Environment and emulated on an FPGA for verification. The intended alternative was a more integrated solution than architectures relying on separate components: a host processor to run the neural network, external memory and a math accelerator.
The announced package was a 15 mm by 15 mm FCBGA324, with support for an industrial operating temperature range. These details described the planned device; at the time of the update, fabrication and engineering-sample availability were still ahead.
Event-based neural networks at the edge
Akida was presented as a spiking neural network (SNN) processor. In this approach, information is represented through events rather than only as conventional, continuously processed values. The company’s update emphasized processing sensor inputs at the point of acquisition, rather than sending them first to a cloud service or data center. For an industrial system designer, the relevant architectural question is whether local processing can meet an application’s requirements for power, response and system integration.
BrainChip cited several power figures for benchmark tasks: 150 µW for keyword spotting using a full network, 200 µW for object detection, and 150 mW for image classification using MobileNet v1. The announcement associated object detection with potential integration into image sensors and described the benchmarks as examples of Akida’s event-processing efficiency. These figures are company-reported benchmarks, not universal estimates for every model, workload or implementation.
The development environment was also intended to support network conversion and simulation. A converter announced in 2019 could translate existing convolutional neural networks (CNNs) into Akida-compatible event-based CNNs, as well as support native SNN translation. That conversion workflow could give developers a way to explore how a network behaves in the Akida environment before targeting hardware.
Licensing the Akida neural fabric
Alongside the chip, BrainChip offered Akida intellectual property for licensing. The neural fabric was organized into nodes, each comprising four Akida cores, with nodes connected in a mesh to support application scaling. The company said the fabric was developed using a standard 28 nm CMOS logic process and could be ported to different process geometries.
Licensing gives an ASIC supplier a different path from adopting a finished device: the supplier can incorporate licensed IP into a custom integrated circuit. BrainChip reported that it was engaging prospective licensees and working with ASIC and design-services providers to broaden access to the IP. It also retained a team for direct IP sales and identified vision and acoustic systems as target areas.
Potential application categories mentioned in the update included automotive radar, LiDAR and cameras for advanced driver-assistance systems and autonomous vehicles, as well as smart cameras, smart-home systems and vision-guided robotics. These were target markets, not confirmed deployments. For engineering teams, the distinction matters: a stated application area indicates where a supplier sees a potential fit, while product selection still depends on project-specific requirements and validation.
Commercialization and adoption timelines
BrainChip’s commercialization plan combined IP licensing with device sales. The company expected licensing to be possible ahead of device sales because licenses would not require a manufacturing cycle and were expected to be prepaid. Device design wins, by contrast, could be established using engineering samples, but customer qualification and design schedules would vary by market.
The update estimated consumer design cycles at six to nine months, while industrial and automotive cycles could extend over several years. This difference has practical implications for planning: a processor evaluation, design-in and production decision may not follow the same timetable across markets. The announcement did not present these timelines as guarantees for individual projects.
BrainChip also described a development agreement with Socionext covering Akida system-on-chip development and manufacturing. The planned wafers were to be produced at TSMC on a 28 nm digital process. The partnership brought together BrainChip’s AI engineering teams and Socionext’s ASIC design experience, according to the company.
What the update means for automation decision-makers
The year-end update presented Akida as both a processor under development and a licensable neural-network architecture. For automation and embedded-system teams, the potential evaluation areas included local vision or acoustic processing, integration options, and the engineering work required to convert or develop networks for an event-based system.
However, the update also made clear that key commercialization steps remained in progress. Fabrication was planned, engineering samples and supporting software were still forthcoming, and adoption schedules would depend on the end market. Any assessment should therefore distinguish between design goals and demonstrated project results, and consider how the intended hardware, software workflow and application requirements fit together.
Conclusion
BrainChip’s 2019 update outlined a two-track strategy for Akida: deliver a neuromorphic processor and license its neural fabric for incorporation into custom ASICs. It identified edge vision and audio as target applications, described tools for CNN-to-event-based-network conversion, and set out the development and design-in stages ahead. For industrial decision-makers, the practical takeaway is to evaluate the architecture and integration path against a specific application while accounting for the maturity and timing of the hardware and development environment.