cudf.pandas: The Zero Code Change GPU Accelerator for Pandas

on YouTube ↗29m

" summary

Claim: run existing pandas code on GPU with zero code changes by swapping in a proxy that routes supported ops to cuDF/CUDA and falls back to pandas when needed. Demo shows big speedups for groupby merge and string ops, a slowdown for count axis=1 due to data movement. Tools: libcudf, importlib import hacks, a notebook profiler that reports GPU vs CPU time. Compatibility notes: 100% pandas API claim with caveats, passes 94% of pandas tests.

AI-written summary · about

" chapters (15)

1. 0:00cudf.pandas introduction
2. 1:05Pandas slowness and alternatives
3. 3:14What is cudf.pandas
4. 5:21Demo setup and usage
5. 6:45Live GPU vs CPU demo
6. 9:35How GPU acceleration works
7. 11:23Third-party library compatibility
8. 12:04Performance optimization tips
9. 13:10Debugging with profiler
10. 15:40Rewriting for GPU efficiency
11. 16:01LangChain acceleration
12. 19:18Architecture deep dive
13. 20:44Pandas compatibility
14. 22:53FAQ about cudf.pandas
15. 24:37Q&A

" related talks

• 30mHacking `import` for speed: how we wrote a GPU accelerator for pandasBradley Dice · PyCon US 2024
• 31mcuDF: RAPIDS GPU-Accelerated Dataframe LibraryMark Harris · PyCon AU 2019
• 29mA Beginners Guide to GPUs for PythonistasMarlene Mhangami · PyConline AU 2021
• 02h 28mGPU Accelerated PythonMcCarty, Riehl, Tomlinson · PyData NYC 2024
• 39mAccelerated Data Science LibrariesAshot Vardanian · PyData Yerevan 2022
• 01h 36mRAPIDS: GPU-Accelerated Data Science for PyData UsersNaty Clementi, Mike McCarty · PyData Virginia 2025
• 01h 12mGPU-Accelerated Data Analytics in PythonKeith Kraus, Bartley Richardson · PyCon US 2020
• 30mCuPy: A NumPy-compatible Library for GPUShohei Hido · PyCon 2018