Renesas Targets the Artificial Intelligence of Things with Its High-Performance Low-Power RA8P1 MCUs

Renesas has announced a new single- or asymmetric dual-core microcontroller range, targeting energy-efficient yet high-performance artificial intelligence and machine learning (AI and ML) at the edge: the RA8P1.

"There is explosive growth in demand for high-performance edge AIoT [Artificial Intelligence of Things] applications," says Renesas' Daryl Khoo in support of the company's latest microcontroller design. "We are thrilled to introduce what we believe are the best MCUs [Microcontroller Units] to address this trend. The RA8P1 devices showcase our technology and market expertise and highlight the strong partnerships we have built across the industry. Customers are eager to employ these new MCUs in multiple AI applications."

The RA8P1 family comes with a primary Arm Cortex-M85 core running at a speedy 1GHz, plus an optional secondary Cortex-M33 core running at up to 250MHz. More importantly for Renesas' vision as the driving force behind an edge AI push is the presence of an Arm Ethos-U55 neural coprocessor, which can deliver up to 256 giga-operations per second (GOPS) of minimum-precision compute at 500MHz — delivering, in real-world terms, up to 35 times the inference performance of the Cortex-M85 core on its own, the company says.

Manufactured at Taiwan Semiconductor (TSMC) on the company's energy-efficient 22nm Ultra-Low Leakage (22ULL) process node, the new chips include either 512kB of 1MB of on-chip magneto-resistive RAM (MRAM) in place of the usual program flash to bring the power draw down still further. Each chip also includes 2MB of static RAM (SRAM) with error correction code (ECC), 32kB instruction and data caches per core, and the option of adding 4MB or 8MB of external flash in a system-in-package (SIP) single-chip sandwich.

Peripherals include a graphics controller supporting outputs up to 1280×800 (WXGA), parallel RGB and MIPI Display Serial Interface (DSI) outputs, a 2D drawing engine, two-lane MIPI Camera Serial Interface 2 (CSI-2) and 16-bit CEU camera inputs, gigabit Ethernet with Time Sensitive Networking (TSN) support, octal-SPI with execute-in-place (XIP) support, SPI, I2C, I3C, SDHI, USB Full Speed and High Speed, CAN FD, PDM and SSI audio interfaces, a digital to analog converter (DAC), 16-bit analog to digital converter (ADC), and a cryptographic subsystem including Arm TrustZone support.

At the same time as the hardware, Renesas has also announced the Renesas Unified Heterogeneous Model Integration (RUHMI) framework. This, the company says, allows for the deployment of neural network models — including native support for TensorFlow Lite, PyTorch, and ONNX — with full optimization for RA8P1 parts. For those looking to get started quickly, the software also includes everything required to deploy pre-trained neural networks including a selection of models and application examples written for the RA8P1.

As is usual for the company, Renesas is offering a range of "Winning Combinations" reference designs that pair the new RA8P1 microcontrollers with the company's other components to build a selection of devices: an AI-enhanced video conferencing camera, an AI-powered robotic drawing arm, and an AI-based surveillance camera.

More information on the RA8P1 chips is available on the official product page; Renesas has also launched the EK-RA8P1 Evaluation Kit development board, which pairs the part with a 7" 1024×600 parallel LCD panel and a five megapixel camera module, priced at $198.