Innatera claims world’s first mass-market neuromorphic microcontroller for the sensor edge
Innatera in the Netherlands has launched the first commercially available microcontroller using a neuromorphic architecture for sensor applications.
The Pulsar chip has a heterogenous architecture that combines analog and digital neuromorphic blocks with a traditional convolutional neural network accelerator and a RISC-V core. This has 100x lower latency and 500x lower energy consumption than conventional AI processors in a 36pin chip that measures 2.6 x 2.8 mm and is built in a standard 28nm process at TSMC for under $5 in volume.
- Innatera shows RISC-V neuromorphic edge AI chip
- €15m for neuromorphic edge AI chip
- Processor designer appoints EDA veteran as chair
“Pulsar is not just another AI chip – this is world’s first mass produced neuromorphic microcontroller and represents a fundamental shift in how we bring intelligence to the edge,” Sumeet Kumar, co-founder and CEO of Innatera tells eeNews Europe.
“38bn sensors shipped last year, growing to 60bn in 2030 and all of these will generate data quicker than we can send it to the cloud, Processing at the edge won’t be optional. But the models that have deployed on microcontrollers have been limited and applications developers have to trade off functions, accuracy and power.”
“This launch is the culmination of over a decade of deep research and engineering in neuromorphic computing, combined with a groundbreaking heterogeneous architecture. It marks the moment that our brain-inspired technology becomes ready for mass-market deployment. This is literally the only microcontroller that a sensor needs.”
The analog neural network (ANN) core uses time voltage pulses to identify patterns and extract information for time series processing without the need for complex models. “The ANN accelerator computes completely with spikes, it’s a large fabric of neurons and synapses with analog and digital devices and a 1ms latency and power under 1mW,” said Kumar.
“Within the fabric the key with a crossbar network of capacitors, and this process is not linear, its exponential, and this is simple with a single transistor in the analog domain,” he said. “The reason why we put in the digital spiking neural network is for configurability and flexibility – that’s done with gates and multipliers. The compute is asynchronous and event driven, and the computations occur at any time when the data comes in. In the CNN you get all the data at one time and compute it.”
“We see a lot of customers with existing ai models that they can just switch over, but generally traditional CNNs look at everything like an image with buffering and all of that takes a lot of power while spiking networks can handle streaming data efficiently. For example a 1m parameter CNN model for gesture recognition can be implemented with a 10,000 parameter, 3kbyte, 54 neuron model with incredibly smaller power consumption.”
“For most application problems you have to pick for AI approach, so adding the CNN developers can add the right tool for the job,” he said.
He points to wireless headsets where the energy of audio sense classification per inference reduced by 100x to 400 µW with a 33x smaller model and same 90%+ accuracy. Sound recognition has 88x lower energy per inference with the same accuracy and latency. Gesture recognition with radar has 42x lower energy than a CNN accelerator at 600 µW and 167x lower latency.
A key part of the chip design is to have interfaces to the sensors, including cameras and medical sensors. Another key element is the software design kit (SDK), called Talamo, and the libraries for the spiking networks.
“The Talamo SDK is built to interface with PyTorch with an extension that brings in all the spiking infrastructure so developers are in a familiar environment, and the model description is in python alongside training data, Our SNN compiler maps the model onto the chip architecture and this radically reduces the barrier to neuromorphic computing, to make it easy to build and deploy spiking models on to frameworks.”
Innatera is launching its developer program, now open to early adopters, with a neuromorphic development board in July. An upcoming open-source PyTorch frontend and marketplace will create an even more collaborative ecosystem for neuromorphic AI.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
