cuDF: RAPIDS GPU-Accelerated Dataframe Library
summary
cuDF is a pandas-like GPU dataframe with a CUDA C++ core and a Python API that delivers concrete performance: microbenchmarks show 30–60× speedups on groupbys/merges/sorts, a 1.9 GB taxi CSV loads in 2 s vs 29 s for pandas, and an end-to-end example reports roughly a 12× cost/speed advantage at cloud-equivalent scale. It supports strings, parquet/JSON/Avro IO, Numba JIT user functions, and scales with Dask and NVLink across GPUs.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:47RAPIDS End-to-End Accelerated GPU Data Science
- 2:38Data Processing Evolution Faster data access, less data movement
- 4:59Data Movement and Transformation The bane of productivity and performance
- 5:47Data Movement and Transformation What if we could keep data on the GPU!
- 7:03Faster Speeds, Real-World Benefits
- 8:49Open Source Data Science Ecosystem Familiar Python APIs
- 9:43GPU-Accelerated ETL
- 11:25ETL: the Backbone of Data Science
- 13:02libcudf Challenges Aim: CUDA-accelerated GPU performance with flexibility
- 15:50Benchmarks: single-GPU Speedup vs. Pandas
- 16:43ETL - the Backbone of Data Science String Support
- 17:39GPU Accelerated 1/0 Pandas API compatibility with 10x speedup
- 20:17CUML Algorithms GPU-accelerated Scikit-Learn
- 23:02Getting RAPIDS
- 23:50RAPIDS Docs
related talks
01h 12mGPU-Accelerated Data Analytics in Python
39mAccelerated Data Science Libraries
29mA Beginners Guide to GPUs for Pythonistas
02h 55mGPU Accelerated Python w: CUDA
29mcudf.pandas: The Zero Code Change GPU Accelerator for Pandas
01h 36mRAPIDS: GPU-Accelerated Data Science for PyData Users
39mAccelerating Data Science with RAPIDS
34mRAPIDS: How to accelerate your data science pipeline by orders of magnitude