cuDF を使ってGPUでDataFrameを処理してみよう

2017:

• Prototyped pygdf, a pandas-like GPU DataFrame • Prototyped dask_gdf, a distributed pygdf

• Ingest data from MapD into PyGDF and do ETL

• Export feature matrices to H2O GLM

2018

• Nvidia created RAPIDS

• pygdf evolved into cudf

• dask_gdf evolved into dask_cudf

https://gyazo.com/4f8884148fd7ef7457fd3cbfb212ba45

Use Arrow format to share data • in-memory

• columnar

• no-copy

• open-standard

• spark, hadoop, parquet, impala, etc..

Using Numba

• Use Numba for GPU functions and GPU arrays • JIT user-defined functions as GPU kernels

• Interactive CUDA programming in notebooks

https://gyazo.com/2486bb323a5b56355528d31fbb82e141

参考