Follow and star our public account for direct access to exciting content
Source: Online materials
Source: Content compiled from eenewseurope.
The Dutch company Innatera has launched the first commercially available microcontroller based on a neuromorphic architecture, suitable for sensor applications.
The Pulsar chip employs a heterogeneous architecture that combines analog and digital neuromorphic modules with traditional convolutional neural network accelerators and RISC-V cores. Compared to traditional AI processors, this chip reduces latency by 100 times and power consumption by 500 times, with a chip size of 2.6 x 2.8 millimeters, manufactured using TSMC's standard 28-nanometer process, with a mass production cost of less than $5.
Sumeet Kumar, co-founder and CEO of Innatera, told eeNews Europe: "Pulsar is not just an AI chip—it's the world's first mass-produced neuromorphic microcontroller, representing a fundamental shift in bringing intelligence to the edge."
Last year, sensor shipments reached 38 billion units, and this number is expected to grow to 60 billion by 2030, with all these sensors generating data faster than we can send it to the cloud. Edge processing will no longer be optional. However, the models currently deployed on microcontrollers are limited, and application developers must make trade-offs between functionality, accuracy, and power consumption.
This release is the culmination of over a decade of in-depth research and engineering in the field of neuromorphic computing, integrating groundbreaking heterogeneous architecture. It signifies that our brain-inspired technology is ready for large-scale market deployment. This is essentially the only microcontroller required for sensors.
The core of the analog neural network (ANN) utilizes time-voltage pulse recognition patterns to extract information for time-series processing without complex models. Kumar stated: "The ANN accelerator fully leverages pulses for computation, consisting of a large number of neurons and synapses, equipped with analog and digital devices, with a latency of only 1ms and power consumption below 1mW."
"In this architecture, the key is a crossbar network made of capacitors, where the process is not linear but exponential, easily achievable in the analog domain with a single transistor," he said. "We added digital pulse neural networks for configurability and flexibility—this can be achieved through logic gates and multipliers. The computation is asynchronous and event-driven, allowing calculations to occur at any time data is input. In contrast, in CNNs, you can only compute after all data is gathered at once."
We see that many existing AI models from customers can be directly switched, but traditional CNNs often treat all data as buffered images, which is very power-hungry, while pulse networks can efficiently handle streaming data. For example, a 1 million parameter CNN model for gesture recognition can be implemented with a model of 10,000 parameters, 3KB, and 54 neurons, with extremely low power consumption.
He noted: "For most application problems, you must choose an AI approach, so increasing the number of CNN developers can add the right tools for the job."
He pointed out that wireless earbuds have reduced inference power consumption for each audio perception classification by 100 times to 400 µW while maintaining over 90% accuracy, and the model size has shrunk by 33 times. The power consumption for each inference in voice recognition has decreased by 88 times while maintaining accuracy and latency. The power consumption for gesture recognition using radar is 42 times lower than that of CNN accelerators (with a power consumption of 600 µW), and latency has decreased by 167 times.
A key part of the chip design is the interface with sensors, including cameras and medical sensors. Another critical element is the software design kit (SDK) called Talamo, along with libraries for pulse networks.
"The Talamo SDK is designed to interact with PyTorch, and its extensions introduce all pulse infrastructure, so developers are in a familiar environment, and the model descriptions are written in Python along with training data. Our SNN compiler maps the models to the chip architecture, fundamentally lowering the barriers to neuromorphic computing, making it easier to build and deploy pulse models on the framework."
Innatera is set to launch its developer program, currently open to early adopters, and will release a neuromorphic development board in July. The upcoming open-source PyTorch frontend and marketplace will create a more collaborative ecosystem for neuromorphic AI.
Reference link
Innatera claims world’s first mass-market neuromorphic microcontroller for the sensor edge
*Disclaimer: This article is original by the author. The content reflects the author's personal views, and the reproduction is only to convey a different perspective, not representing endorsement or support of that view. If there are any objections, please feel free to contact us.
‧‧‧‧‧‧‧‧‧‧‧‧‧‧‧‧ END ‧‧‧‧‧‧‧‧‧‧‧‧‧‧‧‧
Follow the WeChat public account 'ZYNQ' and reply 'join group' to join the FPGA/ZYNQ technology exchange group. Reply 'm' in the background to see more exciting content.