Build a 26 TOPS AI computer from scratch using Raspberry Pi 5 and the Hailo-8 vision accelerator. Every component explained from first principles — what it does, why it was chosen, and how the pieces fit together into a production-grade edge AI platform.
Affiliate links below. If you purchase through them, a small commission is earned at no extra cost to you.
The main compute board — BCM2712 quad-core ARM Cortex-A76 at 2.4GHz with native PCIe 2.0 x1
Splits one PCIe lane into two independent M.2 slots — both drives run simultaneously
Dedicated vision AI chip — object detection, tracking, segmentation at 2.5W typical draw
PCIe Gen 4 x4 storage — OS root filesystem, AI models, recordings. 10× faster than SD card
Mandatory under AI load — Pi 5 will throttle without active cooling. Includes thermal pads
Minimum recommended for Pi 5 with HAT, NVMe, and Hailo-8 all drawing power simultaneously
Compact powered screwdriver — useful for the M.2 retaining screws
Phillips head and flat head for standoffs and case screws
This is a self-contained edge AI computer — a platform that runs vision AI workloads locally, continuously, with no cloud dependency and no ongoing cost. At the heart of it is the Raspberry Pi 5, a credit-card-sized board with a genuine PCIe connector that makes the rest of this stack possible.
The Hailo-8 AI accelerator delivers 26 trillion operations per second of vision AI inference at 2.5 watts. It is a specialist chip — purpose-built for tasks like object detection, people tracking, and multi-stream video analysis. It is not a language model chip. For the applications we build on this channel (security cameras, presence detection, room occupancy), it is exactly the right tool.
The Seeedstudio dual M.2 HAT uses an ASM2806 PCIe switch to split one PCIe lane into two independent M.2 slots. This is what makes it possible to run both the NVMe storage drive and the Hailo-8 accelerator simultaneously from a single Pi 5 PCIe port.
Production architecture. After Video 2, the SD card carries only the bootloader (2 seconds of startup), and the Samsung 990 Pro NVMe carries the full operating system, all applications, and AI models. This is boot-root separation — the same pattern used in production server infrastructure. It provides clean separation of concerns and storage that is 10× faster than a standard SD card.
0:00IntroductionWhat this build produces and why the Hailo-8 was chosen over a GPU-based approach.
0:53The Raspberry Pi 5BCM2712 architecture, the PCIe 2.0 x1 connector, RAM options, and why Pi 5 is the right foundation for this stack.
2:26The Active CoolerWhy passive cooling is insufficient under sustained AI inference load. Push-pin mounting, thermal pad placement.
3:18The Seeed Studio Dual M.2 HATHow the ASM2806 PCIe switch chip splits one lane into two — allowing NVMe storage and the AI accelerator to operate simultaneously.
4:25The Samsung 990 Pro NVMe SSDPCIe Gen 4 x4 performance, capacity, and its role as the OS root filesystem. Why SD card storage is not suitable for production AI work.
5:26The Hailo-8 AI Accelerator26 TOPS at 2.5W typical — what TOPS means, what vision tasks this chip handles, and what it cannot do (LLMs, generative AI).
6:57The 45W Power Supply UnitPower budget calculation — Pi 5 plus HAT plus NVMe plus Hailo-8 under load. Why the official 27W supply is insufficient.
7:40Supporting ComponentsStandoffs, thermal pads, retaining screws, and the microSD card used for the bootloader only.
8:48The Build — Cooler and HATStep-by-step physical assembly: cooler first, then HAT alignment, standoff threading, ribbon cable routing.
13:00HAT and Case AssemblySeating the M.2 drives into the correct slots, securing with retaining screws, final HAT mounting.
23:40Final Connections and Power UpPCIe ribbon cable, USB-C power, first boot confirmation. What to expect on initial power-up.
25:59What's NextVideo 2 preview — headless OS setup, PCIe configuration, NVMe root migration, and Hailo-8 driver installation.
Key details that are easy to miss during a first build. Read these before you start.
The active cooler must be installed before the dual M.2 HAT. Once the HAT is on, you cannot reach the cooler push-pins without removing it. Install in this order: cooler → HAT.
Both M.2 slots on this HAT are electrically equivalent — either device works in either slot. In this build, the Samsung 990 Pro is in slot 1 and the Hailo-8 is in slot 2.
The PCIe ribbon cable is directional. The contacts face down on the Pi 5 connector and face up on the HAT connector. Inserting it reversed will prevent any PCIe device from being detected.
These screws are small and the threads strip easily. Finger-tight plus a quarter turn is sufficient. Do not use power tools on the retaining screws.
After Video 2, the SD card carries the bootloader only. It must remain inserted at all times — remove it and the Pi will not start. This is by design, not a limitation.
Hardware assembled — now make it useful. Flash Bookworm Lite, SSH in with no monitor ever plugged in, configure PCIe Gen 3 for the dual HAT, migrate the entire OS to the NVMe drive, and install the Hailo-8 drivers with a single command. Every step has a copy-paste guide.
View Video 2 guide