AI
GPU-accelerated AI workloads: Ollama, ComfyUI, Kubeflow, and the NVIDIA device plugin.
All AI workloads run on k8s-worker4 (192.168.1.224), which has PCIe passthrough to an NVIDIA RTX 5070 Ti (16 GB VRAM, Blackwell sm_120).
GPU time-slicing advertises 5 virtual GPU resources from 1 physical GPU, allowing Ollama, ComfyUI, and Kubeflow notebooks/training jobs to run concurrently. VRAM is not isolated between processes.
| Page | What it covers |
|---|---|
| Ollama | LLM inference server |
| ComfyUI | Image generation — off by default |
| Kubeflow | Notebooks and pipelines, without Istio |
| Notebook Gateway Controller | Creates HTTPRoutes for notebooks |
| NVIDIA Device Plugin | GPU advertisement and time-slicing |
Hardware details and the VRAM constraint are on GPU.