In this AI-focussed ipXperience, Guy chats with Sumeet Kumar, CEO of Innatera, about artificial intelligence and Innatera’s implementation of this emerging technology in low-power, always-on devices. Prepare for some incredible performance benchmarks…
To begin, Sumeet outlines the origins of Innatera in 2018 as a spin-out from Delft University of Technology. This came after 10 years of research in the fields of neuromorphic and energy-efficient computing. The aim was to make sense of these concepts in electronic hardware and bring the human brain’s ability to find patterns in sensor data to new ‘neuromorphic’ – A.K.A. ‘brain-like’ – processors.
With the requirement for such operations to be performed with highest efficiency and lowest latency, the key use cases were for always-on devices that are constantly taking in data. Innatera now produces semiconductor and software solutions to put this neuromorphic technology into the hands of design engineers.
Guy then asks Sumeet about his experiences of building a semiconductor company in this way, and Sumeet provides some interesting insights into the current state of research in European universities and the initial challenges of the Innatera spin-out process when the team had medical devices in mind. This changed after the focus shifted from applications to the value add of the technology itself – low-power, low-latency processing – which took Innatera into the industrial space, wearables applications, and more.
Sumeet then shows us the evaluation board for Innatera’s latest device, the Spiking Neural Processor T1, a 32-bit, RISC-V neuromorphic microcontroller with a footprint of just 2.16 x 3 mm! The key innovation within this device is Innatera’s ‘analog spiking neuron-synapse array’. This processing architecture provides up to 10,000x higher performance per watt, up to 500x lower energy expenditure, and 100x shorter latency than typical devices trying to perform the same task. As with many AI innovations that ipXchange covered from CES 2024, this is all on-chip processing – no connection to the cloud is required, thus maintaining data privacy.
Additionally, Innatera’s neuromorphic architecture can not only detect patterns in sensor data, but it can also filter the data to make patterns more pronounced – i.e., it can ‘clean’ datasets. This can then be used to train the system for increasing effectiveness in carrying out AI tasks; Sumeet explains this in intricate detail.
Target application workloads include always-on audio processing, radar-driven touch-free interactions, gesture recognition, medical data trend detection – such as for ECGs – and more. The Spiking Neural Processor sits between the sensor and the rest of the system to provide directly actionable inferencing for the fastest response times.
The T1 is currently at the sampling stage, and Innatera expects volume production by the end of 2024. Current test markets include consumer devices, IoT, portable security devices, hearables, smart home, and even automotive, so if you’ve got a commercial project and want to put the T1 to the test, follow the link to the board page below, and apply to evaluate this intriguing technology today.
Keep designing!