Simple GPU Techniques for Big Heliophysics Problems

Simple GPU Techniques for Big Heliophysics Problems Lisa Knowles1, Alexander Antunes1, Nicholas Lenzi1, Peter Shumate1, Sarah Rourke2, Omar Shalaby2, Brian Thomas2, Jeffery Bradford2 [1] JHU APL, [2] NASA GSFC The Problem Heliophysicists want to use accelerated computing to tackle their computationally-demanding research problems, but may lack the resources and training to do so. Our Answer A suite of GPU-accelerated heliophysics Jupyter Notebook tutorials within the HelioCloud platform, derived from published science problems. CPU vs GPU Speed Test Demonstrate the capacity for the GPU to run nearly 200X faster than the CPU GPU Dask Cluster Demo Spin up a cluster of g4dn.xlarge GPU instances HelioML GPU Demo Accelerate HelioML.org’s “Predicting CMEs” Jupyter notebook with cuML-accel Currently Available GPU Tutorials Potential Future Capabilities of HelioCloud • Numba • JAX • cuML-accel • CuPy Future GPU Tutorial Topics • GPU cluster tuning • Intro to PyTorch • Dask-ML • Big data Looking forward... Please take a look at the three prompts below and respond by writing directly on this poster! We want your feedback! 1. On a scale from 1 to 10 (1 being "totally unfamiliar" and 10 being "I do this stuff everyday"), what is your current familiarity with GPU acceleration? 2. Describe your research question/data analysis that could benefit from (or already uses) a GPU-accelerated workflow. 3. Which of the tools described on this poster (or any tool that you know of) do you want to know more about? Answer any of these questions and receive a custom HelioCloud DASH 2025 Challenge Coin! Come find me (Lisa) to claim your reward! HelioCloud GPU servers use AWS g4dn.2xlarge instances, and are equipped with: • 1 NVIDIA T4 Tensor Core GPU • 8 Intel Cascade Lake CPUs • 32 GB RAM