IP-Metis: Edge AI inference
TUM engineering internship project — running vision and communications models on a Raspberry Pi 5 with an Axelera Metis M.2 accelerator, compared reproducibly against the CPU.
For my Ingenieurpraxis (engineering internship) at TUM I brought up an Axelera Metis M.2 AI accelerator on a Raspberry Pi 5 and compared CPU and Metis on several computer vision workloads.
What I did
- Brought up and debugged the accelerator on the Raspberry Pi 5
- Compared computer vision models on CPU and Metis
- Built a 6G-related communications model end to end: training, ONNX export, quantization and compilation for Metis
- Wrote benchmark and evaluation tooling in Python and C++
Now
The project continues as a working-student job, focused on a reproducible CPU / Metis comparison including power measurements. The code is internal to TUM and not public.