While NPUs were established for TinyML in personal and work devices, they have only recently begun to make inroads in IoT applications.
Embedded chipset vendors are increasingly focusing on neural processing units (NPUs) for Internet of Things (IoT) applications, thanks to the architecture’s efficient execution of neural network workloads. As users seek greater insight and intelligence at the far end, NPUs will account for an increasing share of overall shipments at the expense of existing microcontrollers (MCUs). Chipset revenue for AI-specific chips for IoT applications will exceed $7.3B by 2030, according to ABI Research.
“NPUs for TinyML applications in personal and work devices (PWD) are well established. However, they are still in their infancy outside of this device vertical, with major vendors such as STMicroelectronics, Infineon, and NXP Semiconductors only just introducing this type of ASIC into their embedded portfolios,” said Paul Schell, industry analyst at ABI Research. “Screening PWDs provides deeper insights into modeling IoT applications across 15 verticals, including the most important smart home and manufacturing.”
On the software side, comprehensive MLOps toolchains are now a “must have” for vendors large and small, including startups such as Syntiant, GreenWaves, Aspinity, and Innatera. For larger form factors, investments in software products are often matched with hardware R&D, with vendor Eta Compute partnering with NXP to license its Aptos software platform. These innovations also democratize the deployment of TinyML by reducing the need for in-house data science talent.
The incorporation of high-performance architectures, such as NPUs and some FPGAs, into embedded devices will expand the range of applications that can be run on the devices, from object detection to simple object classification for machine vision use cases, and some NLP for audio-based analysis. “With the trend towards larger edge form factors such as PCs and gateways, this will aid in the scalability of AI by reducing network costs and reliance on the cloud. As a result, we expect the TinyML market to grow with these innovations, driven primarily by major industrial bases upgrading their IoT deployments, the increasing intelligence of vehicles, and smart home devices.”
Reposted from: International Electronic Commerce, automatically translated by Google
