Keynote Address: Insights Gained from Delivering Two Generations of AI Supercomputers...
summary
IBM Cloud deploys two generations of Vela systems (2022 → expanded 2023/2024) that virtualize GPUs with ~90% efficiency, move from TCP to GPU direct RDMA over Ethernet to get 2–4× throughput and 6–10× latency gains, and add rack-level power throttling to run 40 kW safely. Workflow handles 10s of PB of raw data down to trillions of tokens (compute ≈ 6 × params × tokens), uses object store + cached file system (Spectrum Scale), and runs an open-source autopilot for continuous GPU health and preflight checks.
AI-written summary · about
chapters (6)
- 0:00Introduction and personal journey
- 1:50Overview of AI advancements
- 9:04AI workloads and infrastructure
- 26:42Vela AI supercomputer systems
- 42:35Vela data architecture
- 49:41Summary and concluding insights
related talks
33mIBM's Approach to Building a Cloud-Native AI Platform
24mClosing Keynote
29mLive Keynote
10mKeynote Analysis
12mKeynote: AI Infrastructure
26mKeynote: Artificial Innovation Accelerates Artificial Intelligence
05mSponsored Keynote: Towards Clouds of AI Clusters
01h 13mEngineering the Cloud for Tomorrow: Performance, Speed and Scale for AI Workloads