Innatera Productizes SNN Accelerator As ‘Neuromorphic Microcontroller’
By
Sally Ward-Foxton 02.06.2024 0
Neuromorphic chip startup Innatera recently productized its spiking neural network accelerator in the form of a “neuromorphic microcontroller,” designed for always-on sensing applications in consumer electronics and the IoT.
The company has become one of a select few with hardware development kits available based on brain-inspired, spiking neural network technology, where information is encoded in the form of precisely timed voltage spikes. SNNs allow native time series processing with very good correlation detection, and they tend to be around 100 times smaller than conventional, non-spiking models, Innatera CEO Sumeet Kumar told EE Times. He said Innatera’s tech has been proven in five generations of test chips, and he added that test silicon validates the company’s claim to achieve 100X the speed and 500X lower energy per inference versus standard neural networks running on digital AI accelerators, including digital signal processors and microcontrollers.
“[The T1] allows you to analyze sensor data in real time to detect and identify patterns of interest through signal processing,” Kumar said. “We tend to focus on applications where there is an always-on sensing element. Generally in this sort of data, events of significance tend to be somewhat sporadic, so we want to do sub-milliwatt processing of this data continuously, and to be able to recognize patterns or carry out signal processing within a millisecond.”
Sumeet Kumar (right) and the Innatera team (Source: Innatera)
Innatera’s new SoC, the T1, features a small CPU, memory and a small convolutional neural network accelerator alongside the company’s analog spiking neural network accelerator.
“Previous prototypes didn’t include a CPU because we were perfecting the neuromorphic fabric itself,” Kumar said. “But at the end of the day, the solution has to be an SoC. We want to be the first chip the sensor talks to, and for most consumer and IoT devices, you need a very small form factor. Being a microcontroller gives the application developer the power to do everything within the chip, so you don’t need any other processing chips close by.”
Innatera‘s T1 chip includes the company’s analog/mixed-signal spiking neural network accelerator (Source: Innatera)
The T1’s CPU is a small RISC-V design with standard sensor interfaces. It performs housekeeping tasks, but it also lets the user configure the sensor to control the flow of data between the sensor and the T1, offloading tasks from the application processor. The CPU can also be used for signal processing pre- or post-inference, if required.
Innatera’s analog/mixed-signal SNN accelerator is programmable such that different SNNs and their complex neuron connection topologies can be programmed onto the chip.
“The array looks a lot like an analog FPGA: You can think of it as being a sea of programmable neurons and synapses that can implement any sort of [spiking] neural network topology,” Kumar said. “The compute within these neurons and synapses is done using analog/mixed-signal computing elements, which on one hand allows great resolution but on the other hand offers unprecedented energy reduction over a conventional digital accelerator architecture.”
SNNs are mapped onto the chip and run continuously. Because SNNs are event-driven, no dynamic power is consumed if no relevant events happen. Compared with previous prototypes, Innatera has further optimized its compute elements for power dissipation, functionality and reliability, Kumar said.
“We’ve arrived at a design point that we’re very happy with, we see very reliable performance, so this was really a process of optimizing and polishing these circuits for the final product,” he said.
Innatera’s T1 includes SNN acclelerator, RISC-V CPU, memory and a CNN accelerator (diagram not to scale – actual area taken up by CNN accelerator is smaller) (Source: Innatera)
Innatera has also added a small digital accelerator on chip for conventional convolutional neural networks (CNNs). This is to ensure the T1 is the only chip needed to turn raw sensor data into actionable insights, Kumar said.
“On our chip, we’re able to do a lot very compellingly with spiking neural networks, but we’ve also demonstrated immense value in coupling the spiking neural network with the CNN for applications that require higher performance, but still within that ultra-low power envelope,” he said. “Having the [CNN] accelerator there gives us additional flexibility to do what’s much harder to do right now with traditional microcontrollers, even those that include some AI acceleration capabilities.”
This allows the T1 to analyze both spatial data (like images) and temporal data (like audio), as well as spatio-temporal data. For example, Innatera’s demos at CES this month included the T1 alongside a radar sensor, with the T1 looking at successive frames of radar for person presence detection or hand-gesture recognition. The demo used a 60-GHz, consumer-grade radar sensor, and the T1 consumed less than a milliwatt of power (less than half a milliwatt for hand gestures) with sub-millisecond latency. Innatera also demonstrated audio scene classification and sound recognition.
Innateral provides an SNN model zoo, and the company’s software stack, Talamo, has a Pytorch front-end including Innatera’s specially-built Pytorch extensions for SNNs.
“The extension adds infrastructure to develop spiking neural networks and train them, and it’s completely built in to the Pytorch framework,” Kumar said. “So if you know how to use Pytorch, you can use our extension to construct these SNNs…. It is a very typical ML-development workflow inside Pytorch to train and optimize models.”
Talamo’s compiler automatically maps Pytorch SNN models onto Innatera hardware without the user needing to know or understand the architecture, Kumar said.
Innatera, a spinout from the University of Delft, has grown to 65 people with recent funding from the European Innovation Council (15.5 million Euro) alongside Matterwave Ventures and MIG Capital.
Commercial samples of the T1 and hardware evaluation kits are available now while the T1 will ramp to production quantities in the second half of this year.