Hey Axelera community! 👋
We're a small optical microscopy engineering company, and we've been exploring whether the Metis M.2 could bring edge AI inference into real research lab environments — not just benchmarks, but actual unattended scientific workflows. Here's what we found.
🧫 What we built
A complete pipeline for real-time single-cell tracking of live Paramecium microorganisms, combining a motorized XY microscopy stage with YOLOv8-seg inference — and ultimately running it on Axelera Metis M.2 + Radxa Rock 5 ITX+.
Real-time motorized tracking with YOLOv8-seg
The core system: YOLOv8-seg detects and segments live Paramecium specimens in real time, driving a motorized XY stage (Nikon/Märzhäuser) to keep individual microorganisms continuously centered in the field of view.
Long-duration trajectory reconstruction
What the system actually captures: a real individual trajectory reconstructed from an extended recording session. We've tracked single specimens continuously for up to 5h 49minutes. This kind of long-duration, unattended data collection is exactly what research labs need for quantitative behaviour experiments.
Metis M.2 proof of concept on Radxa Rock 5 ITX+
This is the edge deployment piece — and where things got interesting. Getting Metis running on the Radxa Rock 5 ITX+ (RK3588) wasn't straightforward, so here are the key steps and gotchas in case it helps anyone:
Steps that actually worked:
- OS: Ubuntu 24.04 Joshua-Riek build (not stock Ubuntu)
- Compiler: Python 3.12 required for us
- Model compatibility: YOLOv8n-seg ✅ — YOLOv11-seg ❌ (fails due to unsupported attention operators at the time that we tried)
- The critical fix — DTB patch: The Radxa Rock 5 ITX+ PCIe MEM window for
fe150000is only 14 MB by default. Metis needs at least 512 MB. We patched the DTB to expand it to 512 MB in 64-bit address space (0x900000000). Without this patch, Metis simply won't initialize. If you're hitting mysterious PCIe failures on ARM SBCs, maybe this is likely your culprit.
Results:
- ⚡ 88.6 FPS with YOLOv8n-seg
- 🖥️ 16.1% CPU usage
- 🌡️ 38°C — thermally stable for 24/7 operation
Why this matters for research labs
The combination of Metis M.2 + an affordable SBC like the Radxa Rock 5 ITX+ opens up a genuinely useful deployment scenario: unattended AI inference in research laboratory conditions. No GPU workstation, no active cooling concerns, no high power draw. Just a compact, stable platform running continuously alongside a motorized microscope.
For frontier science applications — long-duration behavioral studies, automated screening, continuous environmental monitoring — this kind of edge AI node could be a game changer.
Happy to share more technical details on the DTB patch, the YOLOv8-seg training pipeline, or the stage control integration. Just ask! 🙌
Regards,
FxGc
