Syntiant Launches NDP115 Neural Decision Processor for On-Device ML at an Ultra-Low-Power Draw
Offering speech inference at a power draw as low as 280µW, the NDP115 looks to make a home in wearables and smart appliances.
Energy-efficient machine learning specialist Syntiant has announced a new entry in its Neural Decision Processor line-up, the NDP115 ― taking aim at on-device audio and sensor processing in as low a power envelope as possible.
"The NDP115 offers the multi-modal functionality of our Core 2 inference engine in a compact, cost- and power-efficient solution for ultra-power and size constrained applications," claims Kurt Busch, Syntiant chief executive officer, of the new part. "Combined with our machine learning software models, the purpose-built NDP115 enables developers to easily deploy full audio and sensor processing solutions that address all kinds of consumer and commercial use cases, from home security to industrial IoT [Internet of Things]."
Syntiant's original NDP100 family of Neural Decision Processors had shipped over a million units by August 2020, and was joined in January 2021 by the more powerful NDP120 as the first to be built around the Syntiant Core 2 tensor processor platform. The more compact NDP115, as the name implies, sits just below the NDP120 in a 25-ball 0.4mm-pitch WLGBA package to its bigger sibling's 42-ball package — offering the ability to run multiple neural network loads at a power draw of under 1mW and speech inference as low as 280µW.
In addition to the Core 2 neural processor, the NDP115 includes an embedded Arm Cortex-M0 microcontroller, a user-programmable HiFi-3 digital signal processor (DSP) with its own 144kB data RAM and 64kB instruction RAM, an I2C controller, up to 13 general-purpose input/output (GPIO) pins, a security engine for firmware decryption and authentication, and support for direct input of audio across up to five streams.
The company is positioning the new part as ideal for running multiple deep neural networks, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs) on-device with no need to shuffle data off to the cloud, with a particular applicability to "hearables and wearables," smart home, and industrial Internet of Things (IIoT) devices. Those looking for on-device vision capabilities, meanwhile, are still pointed towards the NDP200 launched in September 2021.
Syntiant is showing off the new NDP115 at the Consumer Electronics Show (CES) in Las Vegas this week; pricing has been set at $3.25 per chip in 10,000-unit tray quantities, with immediate availability. Interested parties can book a technology tour on the company's website.