Keynote: From Platforms to AI Factories. Has Kubernetes Solved It?
summary
Claims that Kubernetes plus CNCF tools already run an AI factory: HAMi slices a Blackwell GPU into multiple memory-budgeted devices so one physical GPU serves two open‑source LLMs simultaneously, DRA makes GPUs first‑class devices with topology and device classes so the scheduler grasps rich resources, and vcluster provides tenant isolation atop a shared GPU fleet. Demo uses a MacBook control plane and an NVIDIA DGX Spark; live demo partly fails but a backup recording shows ~95% GPU utilization and working inference.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:06AI factory analogy Mumbai
- 2:08Shift to AI platforms
- 3:09Kubernetes GPU limits
- 4:38Hami GPU virtualization
- 6:22Kubernetes resource labeling
- 7:05Dynamic resource allocation
- 8:09GPU fleet management
- 9:12Demo architecture V cluster
- 13:59Kubernetes for AI factories?
related talks
06mSponsored Keynote: Building Your AI Factory: Wh...
09mKeynote: Kubernetes for GenAI: The New Operating System of the AI-First World
10mKeynote: Enabling AI with Cloud Native- 10mKeynote: Enabling AI with Cloud Native
16mKeynote: An Open-Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLM
32mKeynote: Powering Generative AI with Kubernetes
27mHarnessing Kubernetes for Scalable AI/ML Workloads
20mOperating Frontier Intelligence at Scale