These combinations are present in the first products from Alif Semiconductor Inc. (Pleasanton, Calif.), the Ensemble and Crescendo families of ARM-based AI-enabled, connected microcontrollers (see MCU startup Alif samples two AI-enabled families ).
The company has declined so far to disclose what manufacturing process technology the company is using. More details are set to be released when the microcontrollers go to volume production in 1Q22.
However, using early samples of its devices the company is now able to demonstrate a 76x uplift in power efficiency and a 75x increase in inference performance for some edge node AI/ML workloads using the Ethos-U55 microNPU and Cortex-M55 CPU core. Measurements on Alif's Ensemble MCU show a 76x increase in power efficiency per inference for image classification when compared to using the Cortex-M55 alone.
The demonstration was set up with two subsystems: one was a Cortex-Ethos pair optimized in the implementation for high efficiency (HE) and a maximum clock frequency of 160MHz. A second Cortex-Ethos pair is characterized for high performance (HP) via its implementation and can operate at up to 400MHz.
An audio always-on keyword recognizing demo is run on the first subsystem (using the DN-CNN model) and only when a keyword is recognized is the second pair of cores woken up and an image captured and classification to recognize an object in the frame. This second operation is done using the MobileNetV2 model.
The HP subsystem has additional memory beyond that of the HE system to run more complex decision and classification models. Both the HE and HP subsystems operate on top of Alif’s Autonomous Intelligent Power Management (aiPM™) fabric that ensures power is only consumed by portions of the device that need to be active, extending battery life even further.
In the high-performance subsystem used for image classification the per inference performance of the Ethos-U55 + Cortex-M55 is 75.2x faster and 76.0x more energy efficient than Cortex-M55 alone.
In the high-efficiency