PETaflop Cluster (justingarrison.com)

An enthusiast built a portable two‑node Kubernetes “PETaflop” cluster around an NVIDIA DGX Spark and a LattePanda IOTA board, packed into a wearable backpack to demo local AI inference at KubeCon. The DGX Spark ran ComfyUI for img2img processing while the IOTA hosted the Kubernetes control plane, an ngrok operator for ingress, and an AI‑written frontend that let conference-goers scan a QR code, upload a photo, and receive a stylized image. The project was deliberately constrained to be fully local and portable, showcasing what a compact, edge AI workstation feels like in practice. Technically, the build used off‑the‑shelf parts (GL.iNet travel router, 250W portable power station, mini monitor, 3D‑printed shelving) and a simple k8s split: control/frontend on the IOTA and heavy inference on the Spark to avoid repeatedly reinstalling the Spark. Key operational notes: Spark idled at ~50 W, yielding just over three hours runtime and long recharge times (~2 hours from 30% with the Spark off). The author ran into messy model discovery and broken links in the AI ecosystem, relied on Claude to implement workflows, and found USB tethering more reliable than conference Wi‑Fi. The build underscores practical challenges and tradeoffs for portable local inference—power, connectivity, tooling—and demonstrates a hands‑on path for edge AI demos.