cuDF Guide: GPU DataFrames for Data Science

Introduction

cuDF is a GPU DataFrame library that provides a pandas-like API for GPU-accelerated data manipulation. Part of NVIDIA RAPIDS, it enables 10-100x speedups for data processing.

CUDA Integration: cuDF stores data in GPU memory using Apache Arrow format. Operations are executed as CUDA kernels, with automatic memory management and optional spill-to-host for large datasets.

Key Features

• ✓Pandas-compatible API
• ✓GPU-accelerated operations
• ✓Parquet and CSV I/O
• ✓GroupBy and aggregations
• ✓Joins and merges
• ✓String operations
• ✓DateTime handling
• ✓Dask-cuDF for multi-GPU

Installation

Install via conda.

conda create -n cudf -c rapidsai -c conda-forge -c nvidia cudf=24.02 python=3.10 cuda-version=12.0

conda activate cudf
import cudf; print(cudf.__version__)

Basic Example

Basic DataFrame Operations

Using cuDF like pandas.

import cudf
import pandas as pd

# Create from dict
gdf = cudf.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})

# From pandas
pdf = pd.DataFrame({'x': range(1000)})
gdf = cudf.from_pandas(pdf)

# Operations
result = gdf.groupby('a').mean()
filtered = gdf[gdf['a'] > 1]
merged = gdf.merge(other_gdf, on='key')

Advanced Example

ETL Pipeline

Complete data processing pipeline.

import cudf
import dask_cudf

# Read large dataset
gdf = cudf.read_parquet('data/*.parquet')

# Clean
gdf = gdf.dropna()
gdf['date'] = cudf.to_datetime(gdf['timestamp'])

# Feature engineering
gdf['day_of_week'] = gdf['date'].dt.dayofweek
gdf['rolling_mean'] = gdf.groupby('category')['value'].transform('mean')

# Aggregate
result = gdf.groupby(['category', 'day_of_week']).agg({
    'value': ['mean', 'std', 'count']
})

# To pandas for visualization
pdf = result.to_pandas()

Performance Tips

Common Pitfalls

• •Not all pandas functions supported
• •String operations slower than numeric
• •Large string columns use lots of memory
• •Index operations may differ from pandas
• •Some dtypes not fully supported

Benchmarks

Frequently Asked Questions

Resources

Alternatives

Optimize your cuDF CUDA code with RightNow AI - get real-time performance suggestions and memory analysis.