STMicroelectronics (ST) has unified its tinyML development toolchains into a single stack, ST Edge AI Suite, which will cater for all ST hardware going forward, including ST’s microcontrollers, microprocessors and ML-enabled MEMS sensors.

Remi El-Ouazzane

“Great hardware is only going to achieve its potential if coupled with the right software capabilities,” said Remi El-Ouazzane, president of STMicro’s microcontrollers and digital ICs group. “Rather than individual parts, we need to think of an edge AI system, and this encompasses both the hardware and software needed to get to the right level of performance, cost and power budget for the application.”

ST Edge AI Suite is now the unified AI toolchain for ST hardware including STM32 microcontrollers and microprocessors, STM32 parts with dedicated AI acceleration including STM32N6 and STM32MP2, Stellar automotive MCUs, ST ISPUs (intelligent sensor processing units—MEMS IMUs with DSP), and other ML-enabled MEMS sensors. Tools and libraries developed over the last decade for these different product lines, including NanoEdge AI Studio and STM32Cube.AI, will be incorporated into the new ST Edge AI Suite. ST’s development cloud now includes MEMS products and Stellar MCUs alongside STM32 boards.

ST Edge AI Studio incorporates all tools previously developed by ST for edge AI development and deployment

“There is a need for optimization of scarce resources in an embedded system, be it RAM, Flash or battery capacity,” El-Ouazzane said. “ST has invested heavily in optimization software to compress neural network models, quantize those models and improve their performance, and make sure our customers can get the best from the models they have developed.”

Developers can adopt either a bring-your-own-data or a bring-your-own-model approach, said Marco Cassis, president of ST’s analog, MEMS and sensors group.

“We are keeping things simple: a single entry point, whatever the ST device you are working on,” Cassis said. “Guided navigation to a wide choice of tools, the possibility to use the Suite for any ST devices…and a comprehensive collection of educational resources alongside an active community to help develop skills and implement technology more effectively.”

ST Edge AI Studio Edge AI Core block diagram

A new element, Edge AI Core, is designed to optimize and deploy neural networks on any ST device.

“[Edge AI Core] guarantees the best interaction and performance between software and hardware,” El-Ouazzane said. “It’s a proven technology that has been benchmarked in every [ST] MLPerf Tiny inference benchmark result, and it’s now available on every ST edge-AI–enabled device.”

Demonstrated in ST’s MLPerf Tiny results are ST’s Cortex-M–specific ML libraries (an alternative to Arm’s CMSIS-NN), which will be offered free of charge to ST customers. However, ST will also now offer its ML libraries for other Cortex-M devices from competing suppliers, subject to a special license agreement for deployment and production (i.e., not free of charge).

“We believe that Edge AI Suite is so powerful that we decided to open it as widely as possible, in an effort to accelerate the adoption of edge AI, so customers who would like to run their solution either on a dual platform or only on another ARM Cortex-M device, will be able to do so,” El-Ouazzane said. “This illustrates our commitment to the developer community to help them make edge AI a key feature of their next generation of products.”

A forthcoming release for ST Edge AI Suite will allow developers to compile for Cortex-M–based targets other than ST products. Customers often have a heterogeneous portfolio, El-Ouazzane said, adding that ST doesn’t want its customers to feel locked in.

“We believe it’s a great step forward because it’s essentially allowing our customers to have one unified environment, so they can invest fully into Edge AI Suite and accelerate the adoption of the platform,” he said.

Highlighting deployment and maintenance of deployed ML models as a major challenge to edge AI adoption today, El-Ouazzane said the Edge AI Suite platform is also intended to help customers in this regard. The plan is to take an existing datalogger function from NanoEdgeAI Studio, which handles data capture and annotation, and extend it so it can be used for retraining models with fresh data across all the potential hardware targets.

“We could do more,” he admitted. “At this stage, we have some other ideas that that close the loop between deployment and the continuous retraining over time for tuning of the model accuracy, but it’s not yet [available] in the suite.”

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved