Canonical has published a hands-on guide for running AI models in production on Renesas RZ/V microprocessors using Ubuntu Core 26. The point is to put the DRP-AI accelerator baked into that silicon to work, running neural network inference on the device itself, with no cloud round-trip and no data-center GPU.
The DRP-AI (Dynamically Reconfigurable Processor for AI) is the heart of it. It’s a dedicated block inside the SoC that handles inference, and the guide works with two specific boards: the RZ/V2L and the RZ/V2H EVK. On top of that hardware, Ubuntu Core provides the immutable base system and the snap-based update model, which is where the approach earns its keep.
From an ONNX model to a binary that runs on the board
Nothing gets compiled on the device. All the heavy lifting happens on a host machine or inside a Docker container, and the result is then moved to the board. The compiler is DRP-AI TVM, which uses the EdgeCortix MERA Compiler Framework to translate the model into an instruction set the accelerator understands. You start from a model in ONNX format (the example ships with ResNet18/ResNet50) and end up with optimized Runtime Model Data.
One detail that saves real time: you can cross-compile the whole application for arm64 straight from an amd64 host. No ARM machine needs to be in front of you to prepare the binary, which fits neatly with x86 CI pipelines you already run.
Snapcraft as the delivery mechanism
Once the model is compiled, everything is packaged with Snapcraft into a single snap. Canonical maintains an example repository, rzv_drp-ai_tvm_snap, that bundles Renesas’ upstream sample applications. Inside the snap you get the DRP-AI TVM runtime library (libdrp_tvm_rt.so and its dependencies), the compiled tutorial application (tutorial_app_v2ml), and the pre-compiled model data from the models/ directory.
Installation needs devmode confinement, because the application has to reach the accelerator’s hardware interface:
sudo snap install --devmode rzv-drp-ai-tvm-examples_*.snapAnd to run inference:
rzv-drp-ai-tvm-examples.tutorial-appThe default example processes 640×480 BMP images with the bundled ResNet model. It isn’t a toy demo: it shows the full path of compiling off-device, packaging as a single artifact, and deploying reproducibly on embedded hardware.
Why it matters if you work at the edge
If you’re building devices that run computer vision or image classification on site, this pattern solves two problems at once. The first is model portability: you compile once in your CI and get a snap that installs the same way on every unit. The second is maintenance. Because it ships as a snap on Ubuntu Core, the model and its runtime update through the same channel as the rest of the system, with rollback available if something breaks.
Devmode confinement is the obvious caveat for real production, since it relaxes the snap’s security restrictions to reach the hardware. That’s expected in an example that needs direct DRP-AI access, and it marks where you’d tighten the interfaces and permissions before a serious deployment.
The guide is written by Asa Mirzaieva of Canonical’s Silicon Alliances team and was published on 4 June 2026. It fits the broader work Canonical has been doing to make Ubuntu a solid base for AI workloads, applied here to the smallest end of the chain: embedded silicon.
Source
Based on the original article by Canonical on the Ubuntu blog. All technical details, commands and versions come from that post.