As the ultra-large AI market centered around ChatGPT grows, the installation of AI and machine learning technologies in MCU-level semiconductors at the edge is also very active. Accordingly, global MCU manufacturers are focusing on improving edge AI performance in their products by collaborating with companies with AI capabilities.

Taiwanese semiconductor company Nuvoton Technology Corporation, in collaboration with Skymizer, recently announced that it achieved 'MLPerf Tiny Benchmark' leadership in the Cortex-M4 MCU category with the NuMaker-M467HJ evaluation board and Skymizer's ONNC ML optimization.

The Nuvoton M467 series MCUs use Arm Cortex-M4Fs that operate at a typical 200MHz, and are 67% faster than typical Cortex-M4Fs. This is because the controllers leverage Skymizer’s neural network technology, combined with ML software optimizations, to achieve industry-leading inference performance.

MLCommons is an independent machine learning performance benchmarking collaboration that has established itself as a trusted standard for evaluating ML performance across a wide range of systems. For the MLPerf Tiny benchmark, we focus on practical ML use cases running on embedded systems, such as △virtual wake words △keyword spotting △image classification △audio anomaly detection.

The Nuvoton M467 series consists of a rich set of integrated system functions and peripherals, including △512KB of SRAM △1024KB of flash memory △DSP △FPU △DMA △CAN-FD △I2S △USB △camera interface △encryption accelerator △10/100 Ethernet MAC.

These capabilities make TinyML a great choice for system designers building devices that integrate into applications such as smart home automation, smart cities and infrastructure, light edge AI in IoT, and smart manufacturing.

Skymizer’s ONNC compiler has played a key role in optimizing the machine learning software stack for the M467 series Cortex-M4F, resulting in notable gains in inference speed and overall performance improvements for machine learning applications.

The collaboration between Nuvoton and Skymizer aims to provide cutting-edge solutions for ML in power-efficient embedded systems. Nuvoton said that the combination of hardware excellence and software optimization opens up the possibility of ML-based applications on MCUs.

“These achievements in the Cortex-M4 MCU segment of the MLPerf Tiny benchmark are a testament to the efforts of Nuvoton and Skymizer to push the boundaries of machine learning performance in resource-constrained environments,” Nuvoton added.