
"If you're into Raspberry Pi computers and AI, then chances are that you're already using the AI HAT+ that was released back in 2024. This was a game-changer in terms of offloading AI workloads onto the Pi 5. The idea is that this is a low-cost, data-private way to process edge AI workloads without needing to use third-party cloud-based AI services. Well, today sees that board get a significant upgrade with the release of the AI HAT+ 2."
"The AI HAT+ came in two versions -- a standard 13 TOPS (tera-operations per second) version powered by a Hailo-8L AI accelerator chip, and a 26 TOPS version powered by the Hailo-8 AI accelerator. The new AI HAT+ 2 boosts this performance to 40 TOPS thanks to the use of the Hailo-10H chip backed up by 8GB of onboard RAM."
"This performance boost allows the AI HAT+ 2 to handle large language models (LLMs), vision-language models (VLMs), and other generative AI workloads locally, independent from the Raspberry Pi 5 board. It can also handle vision-based AI workloads such as object recognition and pose estimation at about the same 26 TOPS performance as that of the Hailo-8-powered AI HAT+. Those who have already built projects around the AI HAT+ and want to take advantage of the performance boost that the AI HAT+ 2 has to offer, fear not, because according to Naush Patuck, Senior Principal Engineer at Raspberry Pi, "transitioning to the AI HAT+ 2 is mostly seamless and transparent.""
AI HAT+ 2 upgrades the AI HAT+ platform to 40 TOPS performance using the Hailo-10H accelerator and 8GB onboard RAM. Previous AI HAT+ options offered 13 TOPS (Hailo-8L) and 26 TOPS (Hailo-8). The new board enables local execution of large language models, vision-language models and other generative AI workloads independent of the Raspberry Pi 5. Vision-based tasks like object recognition and pose estimation run at about 26 TOPS, comparable to the Hailo-8 model. Transitioning from earlier AI HAT+ models is reported to be mostly seamless. Hailo provides GitHub resources to support generative AI development on the board.
Read at ZDNET
Unable to calculate read time
Collection
[
|
...
]