The Hailo-10H triples the original AI HAT+ throughput — 40 TOPS of edge NPU on a Pi 5. This is the real setup: PCIe config, the right packages, camera pipelines, and what actually changed from the Hailo-8L.
The AI HAT+ 2 (Hailo-10H) is Raspberry Pi's second-generation AI accelerator HAT. Compared to the original AI HAT+ (Hailo-8L, 13 TOPS), the Hailo-10H delivers 40 TOPS — enough to run larger models at real frame rates or stack multiple inference pipelines simultaneously on a single Pi 5.
| Spec | AI HAT+ (Hailo-8L) | AI HAT+ 2 (Hailo-10H) |
|---|---|---|
| Chip | Hailo-8L | Hailo-10H |
| TOPS | 13 | 40 |
| Form Factor | HAT+ | HAT+ |
| Interface | PCIe Gen 2/3 | PCIe Gen 3 |
| HEF Compatibility | Hailo-8L .hef only | Hailo-10H .hef only |
| Runtime | HailoRT 5.3.0 | HailoRT 5.3.0 |
| YOLOv8s FPS | ~180 | ~500+ |
Hailo-compiled model files (.hef) are not cross-compatible between chips. Hailo-8L .hef files will not load on a Hailo-10H and vice versa. If you're upgrading from an AI Kit or AI HAT+ (Hailo-8L), you need Hailo-10H compiled variants from the Hailo Model Zoo or the hailo-rpi5-examples repo. Check the repo's model download scripts — they auto-select the correct architecture.
Start from a fully updated OS. The Hailo packages in the RPi repository track the firmware, so running on a stale system can cause version mismatches:
The AI HAT+ 2 communicates via PCIe. Unlike the M.2 AI Kit, the AI HAT+ 2 uses the HAT+ PCIe interface which is usually enabled by default in recent RPi OS firmware — but Gen 3 speed still needs to be opted into explicitly for full throughput:
The hailo-h10-all meta-package in the RPi OS repository handles all components. As of recent RPi OS updates, it detects the Hailo device architecture automatically and installs the correct firmware and libraries:
hailortcli command-line toolAfter rebooting, confirm the Hailo-10H is visible and reporting correctly. The device architecture field should read HAILO10H:
You want Device Architecture: HAILO10H and Speed 8GT/s for Gen 3. If you see 5GT/s, you're on Gen 2 — check that dtparam=pciex1_gen=3 is saved in config.txt and you've rebooted.
This is the critical step that differs from the Hailo-8L. You need .hef files compiled specifically for the Hailo-10H architecture. The hailo-rpi5-examples repo provides a download script that handles this:
If you've used a Hailo-8L before, delete those .hef files before running examples on your Hailo-10H. Loading a Hailo-8L model on a Hailo-10H will fail with a cryptic architecture mismatch error. The architecture is baked into the compiled model binary.
With Hailo-10H compiled models in place, the examples run identically to the Hailo-8L — just faster:
The fastest path to live camera inference — no Python needed for the core pipeline. The Hailo-10H works with the same rpicam-apps post-processing descriptors as the Hailo-8L, but you'll see higher FPS:
The Hailo-10H's 40 TOPS headroom opens up workloads that weren't practical on the Hailo-8L. You can run multiple inference tasks in parallel without saturating the NPU:
The hailo-rpi5-examples multi-stream scripts let you push multiple camera feeds through the same Hailo-10H concurrently:
Like the Hailo-8L, Ollama and the Hailo-10H operate in completely separate runtimes. Ollama runs on the RPi 5's CPU cores for LLM inference; the Hailo NPU handles vision tasks. Both run simultaneously on an 8GB Pi 5 without contention — the NPU doesn't consume system RAM for inference (it has dedicated on-chip memory).
Hailo-10H on RPi 5, PCIe Gen 3, HailoRT 5.3.0 — NPU-side inference throughput:
| Model | Task | Resolution | Hailo-10H FPS | vs Hailo-8L |
|---|---|---|---|---|
| YOLOv8s | Detection | 640×640 | ~500 | +178% |
| YOLOv8n | Detection | 640×640 | ~1000+ | +122% |
| YOLOv8s-pose | Pose Estimation | 640×640 | ~370 | +185% |
| YOLOv8s-seg | Segmentation | 640×640 | ~290 | +190% |
| ResNet-50 | Classification | 224×224 | ~1200+ | +167% |
NPU-side numbers from Hailo published benchmarks. End-to-end throughput including camera capture, pre/post-processing, and display is lower — typically 30–90 FPS depending on pipeline. The bottleneck at high NPU speeds shifts to CPU-side pre/post-processing and PCIe transfer overhead.
Quick reference for sourcing Hailo-10H compiled models:
| Source | Hailo-10H .hef? | Notes |
|---|---|---|
| hailo-rpi5-examples download_resources.sh | yes | Auto-selects architecture — safest starting point |
| Hailo Model Zoo (GitHub) | yes | Check releases for hailo10h-specific .hef packages |
| Hailo Developer Zone | yes | Full model catalog, requires free account |
| Hailo-8L .hef files (existing) | no | Architecture mismatch — will fail to load |
| Custom ONNX → Hailo Dataflow Compiler | compile req'd | Target hailo10 architecture in compiler flags |
You're loading a Hailo-8L .hef on a Hailo-10H. Re-run download_resources.sh — it will pull the correct architecture. Or delete old .hef files from resources/ and re-download.
Firmware may not have loaded the Hailo-10H profile. Run sudo apt reinstall hailo-firmware and reboot. If still wrong, check the physical HAT is seated fully on the GPIO header.
Gen 3 not active. Verify dtparam=pciex1_gen=3 is saved in /boot/firmware/config.txt and that you've rebooted after the change. Run sudo lspci -vvv | grep LnkSta to confirm.
Log out and back in to reload your PATH, or run hash -r. The CLI binary installs to /usr/bin/hailortcli.
Add yourself to the hailo group: sudo usermod -aG hailo $USER, then log out and back in.
The Hailo-10H runs hotter than the Hailo-8L at full throughput. Ensure active cooling is in contact with the HAT heatsink. Check temperature with hailortcli monitor — aim to keep below 75°C for sustained inference.
Your answers shape what we write next.
Running a different pipeline? Hit a different error? Let's figure it out.