Accelerating Data Science with RAPIDS
summary
RAPIDS keeps data on the GPU to avoid CPU copy/convert penalties, demonstrating loading ~150–200 GB of CSV (seven years of loan data) into GPU memory, completing ETL steps in seconds and running XGBoost training in about two minutes on a DGX with V100s. Benchmarks show single‑GPU parity with Spark and cluster speedups of roughly 10–25x versus CPUs. The stack wires together Apache Arrow pointers with cuDF, cuML, Dask and a GPU CSV reader so pipelines never leave GPU memory.
AI-written summary · about
chapters (21)
- 0:00RAPIDS and GTC EU intro
- 1:31CPU data science challenges
- 3:17GPU vs CPU workflows
- 6:32Why GPUs outperform CPUs
- 7:36Need for GPU data frames
- 8:17RAPIDS open-source acceleration
- 9:20RAPIDS ecosystem components
- 12:11End-to-end GPU processing
- 13:39RAPIDS architecture deep dive
- 16:47cuDF and GPU features
- 18:50Scaling RAPIDS with Dask
- 22:48RAPIDS machine learning and graph libraries
- 25:21RAPIDS vs XGBoost benchmark
- 27:03Out-of-core data handling
- 28:51Performance and accuracy gains
- 30:05RAPIDS vs scikit-learn speed
- 31:51Getting started with cuDF
- 33:37Live RAPIDS demo
- 36:23RAPIDS resources and Docker
- 37:26RAPIDS future developments
- 39:34Resources, community, hiring
related talks
38mRAPIDS- Open Source GPU Data Science
39mAccelerating Data Science with RAPIDS
34mRAPIDS: How to accelerate your data science pipeline by orders of magnitude
01h 08mGPU Data Science with Rapids
01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUs
01h 12mGPU-Accelerated Data Analytics in Python
20mEnd-to-End Data Science on GPUs
01h 36mRAPIDS: GPU-Accelerated Data Science for PyData Users