Silicon Labs Adds AI To MCUs

Silicon Labs has added a small deep-learning accelerator to its newest wireless microcontrollers, reducing the power required for inference of tiny AI models.

Bryon Moyer

While the rest of the industry has been busy adding deep-learning accelerators (DLAs) to general-purpose processors, Silicon Labs has instead added a tiny one to its new Gecko 2 microcontrollers (MCUs) for ultra-low-power systems. Offering a 78MHz maximum clock speed, the Matrix Vector Processor (MVP) in the EFR32BG24 and EFR32MG24 chips achieves 312 million operations per second (MOPS), targeting small workloads. The two devices recently entered full production.

The only difference between the two products is the wireless protocol. The EFR32BG24 covers Bluetooth, and the EFR32MG24 (M for multi-protocol) covers Bluetooth Low Energy, Zigbee, and Thread. In discussions unrelated to wireless, the company refers to the two together as the EFR32xG24, or simply the xG24.

Silicon Labs has two wireless-MCU families: Gecko 1 and Gecko 2. Both execute AI on the CPU using TensorFlow Lite for Microcontrollers; only the xG24 model has MVP assistance.

The MVP is the sole DLA in its class, as most accelerators target trillions of operations (TOPS)—or at least billions (GOPS). The tradeoff is power: the xG24 targets battery-powered IoT systems, where every electron counts. By performing AI locally, the chip can communicate results only if warranted rather than shipping all data to the cloud for analysis (regardless of whether eventual results are interesting) or completely forgoing AI. Expending some energy on AI can reduce energy spent communicating.

The MVP’s inference times are similar to those of designs that run equivalent CPUs at more than twice the xG24’s clock speed. But the big power savings enables the xG24 to outperform the power and efficiency of other MCUs on the MLPerf Tiny models.