Edge AI Inference: A Deep Dive from Software to Hardware Acceleration

2025-07-04
Edge AI Inference: A Deep Dive from Software to Hardware Acceleration

This article delves into the challenges and opportunities of running AI inference on resource-constrained microcontrollers. Starting with the mechanics of TensorFlow Lite Micro, the author analyzes the software implementation and hardware acceleration schemes based on ARM architecture extensions for the addition operator. The article also covers utilizing Arm's Ethos-U NPU for model acceleration. It reveals how different hardware architectures impact AI inference performance and how software and hardware optimizations can be combined to improve efficiency.

Read more

Unlocking the RAK5010's BG95-M3 USB Serial Ports

2025-02-22
Unlocking the RAK5010's BG95-M3 USB Serial Ports

This post details accessing the three serial ports on the Quectel BG95-M3 module embedded in the RAK5010 development board. The author, after consulting the schematic, identified and resoldered four resistors to redirect the USB data lines from the nRF52840 to the BG95-M3. Despite a minor mishap involving a lost resistor, the modification was successful, allowing direct USB communication with the modem and access to its three serial ports.

Read more
Hardware

Barely Running Wi-Fi Station Mode on a Resource-Constrained MCU with Thingy:91 X

2025-02-21
Barely Running Wi-Fi Station Mode on a Resource-Constrained MCU with Thingy:91 X

This post details the author's experience running Wi-Fi station mode on the resource-constrained Nordic Semiconductor nRF9151 MCU using the Thingy:91 X. The nRF9151's limited 256KB of RAM presented a significant challenge due to the Wi-Fi driver's resource demands. By disabling nrf_modem_lib and cleverly using overlay files and devicetree configuration, Wi-Fi connection was achieved, albeit at a whopping 99.11% RAM utilization. The post meticulously analyzes encountered issues, including driver resource consumption, patch loading methods, and socket creation, providing solutions and valuable insights for embedded developers.

Read more
Development

Nordic Unveils VPR: Its First RISC-V Processor, Ushering in a New Era of Heterogeneous Computing

2024-12-26
Nordic Unveils VPR: Its First RISC-V Processor, Ushering in a New Era of Heterogeneous Computing

Nordic Semiconductor has launched VPR, its first RISC-V processor, integrated into the new nRF54H and nRF54L SoCs. VPR, an RV32EMC processor running at up to 320MHz, is designed for software-defined peripherals. The article details VPR's architecture, initialization process, and collaboration with the Arm Cortex-M33. Zephyr's sysbuild simplifies building and deploying VPR applications, enabling heterogeneous computing for enhanced performance and functionality.

Read more