A Hands-On Workshop by Empire AI and NVIDIA
Duration: Approximately 3 hours | Format: Instructor-led instruction,
hands-on single-node labs on NVIDIA A100 GPUs, and demonstrations
Get more from your GPUs by learning how to measure, understand, and improve AI application performance. Join Empire AI and NVIDIA for a practical workshop on profiling PyTorch workloads with NVIDIA Nsight Systems and applying targeted optimizations.
Through selected exercises adapted from the Open Hackathons Profiling AI Software Bootcamp, participants will establish a baseline, capture a performance profile, identify bottlenecks, apply changes, and measure the results. Hands-on labs focus on GPU optimization within a single compute node. Instructor demonstrations introduce multi-node profiling and transformer optimization with Transformer Engine and FP8, using results from compatible hardware.
What you will learn
- Understand how CPU–GPU connections and GPU topology influence application performance.
- Capture and interpret Nsight Systems traces, using NVTX annotations to connect application code with timeline activity.
- Identify bottlenecks in data loading, CPU–GPU transfers, GPU utilization, and communication between GPUs within a node.
- Apply targeted optimizations to a PyTorch training workload and compare performance before and after changes.
- Explore multi-node profiling and Transformer Engine/FP8 optimization through instructor demonstrations.
Three hour agenda
Duration | Activity |
20 minutes | Performance foundations: Profiling workflow and single-node system topology |
40 minutes | Hands-on lab: Capture a baseline profile and interpret the Nsight Systems timeline |
60 minutes | Hands-on lab: Identify bottlenecks and optimize a PyTorch workload on GPUs within one node |
10 minutes | Break |
20 minutes | Demo: Multi-node training, communication, and profiling |
20 minutes | Demo: Transformer optimization with Transformer Engine and FP8, featuring results from compatible hardware |
10 minutes | Review results, practical next steps, and Q&A |
Prerequisites
Working knowledge of Python, familiarity with PyTorch training workflows, and basic comfort with Linux commands and Jupyter notebooks. Prior experience with profiling tools is optional.
Optional preparation videos: CUDA Developer Tools playlist on YouTube. Review ahead of the workshop to become familiar with NVIDIA developer tools. Viewing is optional and is not required to participate.
Leave with a practical approach to finding performance bottlenecks and evaluating optimizations in your own AI workloads.
Was this article helpful?
That’s Great!
Thank you for your feedback
Sorry! We couldn't be helpful
Thank you for your feedback
Feedback sent
We appreciate your effort and will try to fix the article