Accelerating Data Science with RAPIDS
summary
Claims GPU-accelerated end-to-end pipelines using cuDF and cuML with Apache Arrow memory sharing, Dask for distribution, and CUDA-backed JIT for Python UDFs. Demonstrates single-node 16 V100 performance beating a 100-node Spark cluster on an ETL+XGBoost workflow, CSV reads >10x faster than pandas, PageRank speedups reported up to 10,000x (worst-case ~130x), and containerized Conda/docker installs; notes multi-node communication bottlenecks and ongoing work on parquet/orc readers and a CUDA array interface.
AI-written summary · about
chapters (25)
- 0:00RAPIDS introduction
- 1:48CPU limits in data science
- 3:14Pipeline acceleration need
- 4:38GPU adoption barriers
- 6:04GPU data retention
- 7:28RAPIDS libraries overview
- 8:57RAPIDS with CUDA/Python
- 10:20RAPIDS installers/containers
- 11:01cuML machine learning
- 12:04Pipeline performance gains
- 13:31Scaling RAPIDS
- 14:57Algorithm benchmarks
- 15:19New algorithms/features
- 16:43cuDF data transformation
- 19:31cuDF architecture
- 21:15cuDF language design
- 22:47cuDF with Dask/UCX
- 26:17Distributed ML with RAPIDS
- 27:43cuDF improvements
- 28:24GPU string operations
- 29:08GPU data loading
- 30:53Python CUDA Array Interface
- 32:39Open-source contributions
- 33:21Careers at NVIDIA
- 33:51Q&A
related talks
41mAccelerating Data Science with RAPIDS
34mRAPIDS: How to accelerate your data science pipeline by orders of magnitude
01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUs
38mRAPIDS- Open Source GPU Data Science
01h 08mGPU Data Science with Rapids
01h 12mGPU-Accelerated Data Analytics in Python
20mEnd-to-End Data Science on GPUs
01h 36mRAPIDS: GPU-Accelerated Data Science for PyData Users