Raspberry Pi Targets On-Device Edge AI with Its 26 TOPS AI HAT+ Add-On

Raspberry Pi has announced a new add-on for Raspberry Pi 5 users looking to experiment with on-device machine learning and artificial intelligence (ML and AI) workloads: the Raspberry Pi AI HAT+, developed in partnership with Hailo and offering up to 26 tera-operations per second (TOPS) of compute at INT8 precision.

"The Raspberry Pi AI HAT+ is an add-on board with a built-in Hailo AI accelerator for Raspberry Pi 5," the company says of its latest hardware launch. "It provides an accessible, cost-effective, and power-efficient way to integrate high-performance AI. Explore applications including process control, security, home automation, and robotics."

The Raspberry Pi AI HAT+ follows the launch of the Raspberry Pi AI Kit back in June. Developed, like the new AI HAT+, in partnership with low-power on-device AI specialist Hailo, the AI Kit bundled a Hailo-8L M.2 accelerator board with a Raspberry Pi M.2 HAT+ — more commonly used to connect high-speed Non-Volatile Memory Express (NVMe) drives to the Raspberry Pi 5. The new AI HAT+, though, is a fully-integrated design: the accelerator chip is soldered directly on the board, rather than using an M.2 module and carrier.

There's also been a boost in performance: the base model of Raspberry Pi AI HAT+ uses the same Hailo-8L chip as the original Raspberry Pi AI Kit, offering 13 tera-operations per second (TOPS) of INT8-precision compute; the top-end model, meanwhile, moves to the Hailo-8, which offers twice the performance at 26 TOPS.

Installation is easy: the HAT+ board sits atop of the Raspberry Pi 5, using spacers that provide enough room for a Raspberry Pi Active Cooler underneath, and connects to the 40-pin general-purpose input/output (GPIO) header with a pass-through connector and to the Raspberry Pi 5's PCI Express lane via a flat flexible circuit (FFC). This is used for the data connection, giving the AI HAT+ a high-speed PCI Express Gen. 3 lane for maximum throughput — a step above the usual Gen. 2 for which the PCIe lane is rated, though Raspberry Pi tells us it's fully tested at the higher throughput.

Software for the AI HAT+ has already been integrated into the latest Raspberry Pi OS Linux distribution, the company says: once connected, the accelerator is detected automatically and made available to the system for compute acceleration. For computer vision work, support has been added to the Raspberry Pi camera software stack — allowing the accelerator to be used for post-processing work on any compatible Camera Module, rather than just the new Raspberry Pi AI Camera Module with its built-in accelerator.

The Raspberry Pi AI HAT+ is available to order from the company's official resellers now, priced at $70 for the Hailo-8L 13 TOPS variant and $110 for the Hailo-8 26 TOPS model.