IBM's Approach to Building a Cloud-Native AI Platform
summary
IBM and Red Hat describe an open source cloud‑native stack that combines KubeRay & Ray, PyTorch (FSDP for distributed training), Triton CUDA‑free kernels, and OpenShift AI to run end‑to‑end foundation model workflows. The platform runs in production for 2.5+ years on IBM Vela, scaling from single GPUs to thousands of GPUs, uses an “instant slice” project for GPU fractionalization, reports ~30% throughput gains from fully sharded data parallel, and helped train a geospatial model with NASA on 2.5 trillion tokens.
AI-written summary · about
chapters (9)
- 0:00Introduction and speaker intro
- 1:52Cloud-native AI platform complexity
- 3:17IBM's open-source AI stack overview
- 8:56Open-source contributions and integrations
- 13:36Internal platform deployment at IBM
- 15:20OpenShift AI as enterprise support layer
- 17:08Data Prep Kit use case
- 21:25NASA geospatial model training
- 28:29Q&A session
related talks
52mKeynote Address: Insights Gained from Delivering Two Generations of AI Supercomputers...
04mA Cloud Native Stack From Bare Metal To Tokens for Large-Scale AI Inference
35mBuilding Enterprise Solutions in the AI-Native Era
25mEnabling your business for AI
06mTalking Open Source & Cloud Native
25mAI In Production Key Challenges and Use Cases
35mOperating a Production-Scale Physical AI Data Platform with Ray
05mOpen Source Infrastructure for the AI Native Era