Run Your Own AI Cluster on a DGX Spark: Kubernetes, GPUs, and DRA
summary
Start from a single-node Ubuntu-based cluster on a DGX Spark and enable GPUs with either the NVIDIA GPU Operator or the modern DRA driver; DRA runs as Kubernetes primitives (available in recent servers since 1.34) and exposes ResourceClasses and ResourceClaims so the scheduler can match slices of a GB10 GPU rather than just counting devices. Demo shows a GB10 with 128 GB unified memory and ~1 petaflop running vLLM serving Gemma 4 (31–32B params) offline, converting hand‑drawn sketches into YAML and scaling deployments from 4 to 8 pods with live enforcement.
AI-written summary · about
chapters (19)
- 0:00Running AI clusters on DGX Spark
- 1:05Laptop/cloud GPU limits
- 2:10DGX Spark hardware
- 3:13Kubernetes setup on DGX
- 3:53Making GPUs visible to K8s
- 4:55Why Dynamic Resource Allocation
- 6:42DRA vs storage in K8s
- 9:53DRA architecture
- 11:40Live demo setup
- 13:06Verify GPU/cluster status
- 14:49Device class and resource slice
- 15:31Resource claim templates
- 16:14Deploy VLLM with Gemma 4
- 17:18Sketch-to-YAML inference
- 18:43Validate and deploy YAML
- 20:08Scaling with AI YAML
- 21:35Local AI future
- 23:18DRA portability
- 24:43Q&A on DRA, MIG, VRAM
related talks
25mEmpowering AI-Driven Drug Discovery: Overcoming Challenges in Building a ML Platform on Kubernetes
28mA Deep Dive on How To Leverage the NVIDIA GB200 for Ultra-Fast...
32mScaling Kubernetes Clusters for Generative Models: Managing GPU Resources for AI App...
01h 15mTutorial: Build, Operate, and Use a Multi-Tenant AI Cluster Based Entirely on Open Source
16mAI + Kubernetes: What Beginners...
32mLeveraging DRA for Conquering ML Workload Challenges on...
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
40mDemystifying AI Hardware: NVIDIA Evolution and Costs