When AI Meets High Performance Computing
summary
Practical guidance on running deep learning workloads on HPC clusters using PyTorch and Jupyter frontends with Slurm job scheduling and container runtimes like Apptainer/Singularity on A100 100GB GPUs. Covers handling no‑internet compute nodes and predownloaded datasets, tuning batch size per GPU, number_of_workers and pin_memory for CPU→GPU pipelining, and combining data and model parallelism when a model exceeds one GPU; reports measured speedups (×6 from batch‑size tuning on one A100 and ×42 from multi‑GPU scaling on Calmip).
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:17CalmIP and HPC role
- 12:09Deep learning basics
- 21:35Data and model parallelism
- 29:49Multi-GPU HPC code
- 38:41Performance and memory
- 41:54HPC environment access
- 44:25Optimization results demo
- 45:28CalmIP 2026 and conclusion
related talks
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
03mAccelerate AI adoption
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
41mBuilding & Operating a Large-scale HPC AI Cluster on Kubernetes
13mFrom Silicon to Scale: Training at Scale
09m⚡ Lightning Talk: Advanced GPU-Orchestrated Workflows and HPC Integrations on K8s for Distributed AI/ML at Scale
12mKeynote: AI Infrastructure
22mAI @Scale