AI experiments increasingly depend on more than a model and a container. The accelerator underneath them brings kernels, drivers, firmware and hardware-specific operating-system requirements into the reproducibility boundary.

imec’s EdgeLab gives researchers access to a deliberately non-uniform pool of 10+ edge AI device types, including NVIDIA Jetsons. At build time, the system environment defines what each device needs: kernel, accelerator drivers, userspace and other hardware-specific requirements. At runtime, researchers request that hardware through pods on a shared Kubernetes cluster, run their experiments, and can discard and recreate the workload when it breaks.

That separation matters because these devices aren’t meant to stay in the lab. Edge hardware may sit in a datacenter while researchers experiment with it, but eventually workloads need to run in the field. If the system underneath the workload changes along the way, the environment being deployed is no longer the environment that was tested.

Containers solve only part of that problem. Below the kernel, edge hardware does not behave like a homogeneous datacenter fleet. BMCs and network boot cannot be assumed. Jetsons have their own flashing requirements. Different accelerators can require different kernels, drivers and userspace. Kubernetes begins once that machine exists; managing how it got there is a separate lifecycle.

This talk examines the resulting reproducibility problem. In EdgeLab, the reproducible unit extends beyond the container to include the system image that makes the accelerator usable. We will show how image-based machine management provides one operational model while still allowing hardware-specific images to differ deliberately, and how carrying that environment from lab to field can reduce the gap between what was tested and what is ultimately deployed.

Jasper De Keukelaere will show why imec built EdgeLab, how researchers use it, and what the lab can reproduce today. Mauro Morales will examine the machine lifecycle underneath it: system images, hardware-specific divergence, upgrades and rollback.

Finally, we’ll draw the boundary explicitly: which parts of an AI experiment can travel from lab to field today, which remain outside that boundary, and why knowing the difference matters when trying to reproduce an inference result.