CuDF – Maybe faster Pandas on the GPU via RAPIDS (NVIDIA)

Slide 1

Slide 1 text

CuDF – Maybe faster Pandas on the GPU via RAPIDS (NVIDIA) PyDataLondon 2024-02 lightning talk @IanOzsvald – ianozsvald.com

Slide 2

Slide 2 text

NVIDIA tech built on CUDA – “100% Pandas compatible” GPU accelerat. Run your dataframes in VRAM Linux/Windows (WSL) only Will add to 3rd edition of our book + course-> CuDF Pandas GPU Accelerator By [ian]@ianozsvald[.com] Ian Ozsvald

Slide 3

Slide 3 text

Benchmark-winning? By [ian]@ianozsvald[.com] Ian Ozsvald Very hardware dependent – h2oai site has much weaker GPU and smaller machine

Slide 4

Slide 4 text

Benchmark-winning? By [ian]@ianozsvald[.com] Ian Ozsvald

Slide 5

Slide 5 text

RTX 3050 Ti (Laptop) 4GB VRAM vs 16 core i7 64GB Demo 10M rows 3 cols of ints <1GB df By [ian]@ianozsvald[.com] Ian Ozsvald On RTX 3090 (24GB VRAM) I get a 30x speed-up with 400M rows (9.6GB df)

Slide 6

Slide 6 text

“Potential” for 10x-100x faster in production You need >>2X VRAM than your df -> Fiddly to setup, dev experience tricky – suspend, drvr updates, Pandas 2 2024? Let’s discuss in the break Summary By [ian]@ianozsvald[.com] Ian Ozsvald

Slide 7

Slide 7 text

Let’s discuss in the break – what are you building? Summary By [ian]@ianozsvald[.com] Ian Ozsvald Pandas 1.5

Slide 8

Slide 8 text

/home/ian/workspace/teaching/fast_pandas/scratch/pandas_cudf_202312 sudo modprobe -r nvidia_uvm; sudo modprobe nvidia_uvm # maybe solves cudeaErrorUnknown after a suspend Appendix By [ian]@ianozsvald[.com] Ian Ozsvald