GPU-Accelerated Data Analytics in Python

on YouTube ↗01h 12m

" summary

The RAPIDS stack demonstrates concrete speedups: GPU dataframes via cuDF read a 1.9 GB NYC taxi CSV in ~2s versus ~29s with pandas; a 200 GB end‑to‑end pipeline drops from 322s to 209s across releases; DBSCAN on 10k rows runs ~29.7 ms on GPU vs ~31.2 s on CPU; UMAP is ~9× faster; forest inference gives ~2 ms vs ~15.7 ms for XGBoost predictions; PageRank on a 16B‑edge graph runs in ~31.8 s on 16 V100s vs 96 minutes on 108 vCPUs. Demos and a runnable container (cuDF cuML cuGraph cuSignal cuSpatial) are provided and reproduce these results.

AI-written summary · about

" chapters (27)

1. 0:00Intro
2. 0:13RAPIDS GPU Accelerated Data Analytics in Python
3. 2:25Data Processing Evolution Faster data access, less data movement
4. 3:45Data Movement and Transformation What if we could keep data on the GPU?
5. 4:33Faster Speeds, Real-World Benefits
6. 6:22Speed, Ease of use, and Iteration The Way to Win a Data Science
7. 8:13Open Source Data Science Ecosystem Familiar Python APIs
8. 9:00RAPIDS End-to-End Accelerated GPU Data Science
9. 12:29RAPIDS GPU Accelerated data wrangling and feature engineering
10. 13:50ETL Technology Stack
11. 15:55Benchmarks: single-GPU Speedup vs. Pandas
12. 16:26ETL: the Backbone of Data Science String Support
13. 17:40Extraction is the Cornerstone CUDF I/O for Faster Data Loading
14. 19:55ML Technology Stack
15. 21:30RAPIDS matches common Python APIs
16. 33:23RAPIDS RELEASE SELECTOR
17. 37:35Forest Inference Taking models from training to production
18. 40:25Goals and Benefits of cuGraph Focus on Features and User Experience
19. 41:59Graph Technology Stack
20. 42:51Algorithms GPU accelerated Network
21. 45:31Multi-GPU PageRank Performance PageRank portion of the HiBench benchmark suite
22. 52:37cuSpatial Technology Stack
23. 55:08Speed of Light Performance - V100
24. 55:48Efficient Memory Handling
25. 56:18Marriage of Deep Learning and RF Data
26. 59:07CLX Cyber Log Accelerators
27. 59:44CLX Components Notebook Examples, SIEM Integrations, Workflows, Primitives

" related talks

• 41mAccelerating Data Science with RAPIDSMike Wendt · PyData Los Angeles 2018
• 38mRAPIDS- Open Source GPU Data ScienceBrad Rees · PyData Los Angeles 2019
• 01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUsKeith Kraus · PyData New York City 2019
• 39mAccelerating Data Science with RAPIDSKeith Kraus · PyData DC 2018
• 34mRAPIDS: How to accelerate your data science pipeline by orders of magnitudeAdam Grzywaczewski · PyData Eindhoven 2020
• 01h 36mRAPIDS: GPU-Accelerated Data Science for PyData UsersNaty Clementi, Mike McCarty · PyData Virginia 2025
• 01h 08mGPU Data Science with RapidsKeith Kraus, Joshua Patterson · PyData Miami 2019
• 20mEnd-to-End Data Science on GPUsJohn Zedlewski · OpML '20