Differential Software Solutions on HPC and IBM Power Architecture
summary
Claim: IBM Power9 nodes run distributed in‑memory AI training at supercomputer scale using RDMA/Infiniband and NVLink, not just CPUs. Demo builds a Cloud Pak for Data pipeline on Red Hat OpenShift, ingests a small Excel diabetes dataset, trains and deploys models across Power9 nodes and NVIDIA accelerators. Reported result: reproduced a large training run in about seven hours with ~95% efficiency and scale across racks (Barcelona MareNostrum reaches ~1.5 petaflops using this stack). Tools: Cloud Pak for Data, OpenShift.
AI-written summary · about
chapters (18)
- 0:00Introduction to talk
- 1:07Open Power architecture overview
- 4:24Open Power in HPC today
- 8:07Power servers security
- 10:15Power processors performance
- 13:15Supercomputing with Power
- 16:25Power servers components
- 18:15Power9 processor details
- 21:27AI workloads on Power
- 24:03AI training models
- 27:00Open source software solutions
- 28:27AI cloud software stack
- 31:43OpenShift container demo
- 33:08AI model prediction demo
- 37:28Deployment to production
- 39:14Online AI application demo
- 40:43Data science insights
- 41:27Q&A and closing
related talks
25mHAL: Computer System for Scalable Deep Learning
31mHigh-Performance MPI and Deep Learning on OpenPOWER Platform
27mThe convergence of machine learning and high performance computing
25mBenchmarking Graphical Models on POWER with Healthcare...
33mIBM's Approach to Building a Cloud-Native AI Platform
25mHigh-Performance MPI and Deep Learning with Introspection...
44mArchitecture Effects When Interacting with GPU Technologies
40mAI Factory - Enabling Scalable Innovation & Intelligent Systems