Building at Scale with H100: Eos as a DGX SuperPOD Reference Model for Large Data Center Builds
summary
Eos is a DGX SuperPOD reference design built from DGX H100 servers organized into self‑contained pods that scale to ~4,000 GPUs and achieved a top‑500 rank (#9) and commercial MLPerf availability. It uses a leaf‑spine split‑core fabric (Infiniband) for collectives, a separate parallel storage fabric with NVMe caching, liquid cooling via CDUs, Prometheus/Splunk telemetry, Slurm scheduling and validated NVIDIA container images to guarantee performance and incremental deployability.
AI-written summary · about
chapters (8)
- 0:00Building at scale with H100
- 1:52DGX SuperPOD timeline
- 3:58EOS SuperPOD architecture
- 7:27Compute fabric optimization
- 10:18GPU storage architecture
- 12:08Cluster software & telemetry
- 15:59Containerized job scheduling
- 17:03Next-gen DGX challenges
related talks
25mHigh Performance Networking for Distributed DL Training in Production K8s
40mAI Factory - Enabling Scalable Innovation & Intelligent Systems
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
25mHAL: Computer System for Scalable Deep Learning
35mState of Containers and the Convergence of Hyperscale, BigData and HPC
28mA Deep Dive on How To Leverage the NVIDIA GB200 for Ultra-Fast...
31mShared GPU Scheduling & Proactive Autoscaling: A Production Blueprint for 1000+…
25mBuilding High-Performance, Sustainable AI Clusters