A comparison of most commonly used Data-Science Python packages and their alternatives. Generally, those alternatives have identical Python interfaces, but come with Multi-Threaded CPU or even GPU backends, implemented in C++, CUDA, Rust and other low-level languages.

To run the default configuration for every folder - the procedure is similar:

cd x
conda env create -f env.yml
python bench.py

For Linear Algebra and Digital Signal Processing we synthetically generate square random matrices, mainly of with single-precision floating point numbers. That is different from the default Pythons float that uses the 64-bit representation, more commonly described as double in C-like languages. Participating packages:

• NumPy over BLIS
• NumPy over OpenBLAS
• NumPy over Intel MKL and One API
• CuPy over CuBLAS

For Graph Theoretical and Network Science workloads we pick various commonly used datasets from the Stanford Network Repository. All ranging under 1 MB to over 1 GB and 100 million edges. Participating packages:

• NetworkX
• RetworkX
• CuGraph

We took the NYC Taxi Rides dataset as our primary dataset and run the classical 4-query benchmark on its subsets. Participating packages:

• Pandas
• Modin
• CuDF
• Dask-CuDF
• SQLite
• Apache DataFusion