HAL: Computer System for Scalable Deep Learning
summary
Describes a deployed deep-learning cluster built from 16 IBM POWER9 servers (64 NVIDIA V100 GPUs total) connected by Mellanox EDR InfiniBand and a DDN all‑flash array to run IBM Watson Machine Learning alongside open‑source frameworks. Explains a custom management stack, notebook/web interfaces and job recipes that enable full‑GPU packing and multi‑node distributed training, and presents implementation details, performance numbers and step‑by‑step recipes showing PyTorch and TensorFlow models scaling across all 64 GPUs.
AI-written summary · about
chapters (5)
- 0:00Introduction to HAL system
- 8:15Deep learning support overview
- 14:07User interfaces and projects
- 20:34GPS and system applications
- 23:48Q&A and conclusion
related talks
42mDifferential Software Solutions on HPC and IBM Power Architecture
31mHigh-Performance MPI and Deep Learning on OpenPOWER Platform
43mAI APPLICATIONS ON POWER9 SYSTEM
49mDistributed Deep Learning
34mDeep Learning Support at Indiana University
13mFrom Silicon to Scale: Training at Scale
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights