Accelerated Data Science Libraries
summary
Replaces Pandas/NumPy/NetworkX with concrete backends and shows measured outcomes: switching to cuDF or cuPy yields tabular 10–50x speedups and math kernels sometimes reach ~1000x while some cases become slower; Modin falls behind cuDF on sampled New York taxi benchmarks; GPU memory limits, dtype/casting differences, and explicit synchronization often force small code changes (e.g. import cupy as np) and tuning; a working persistent engine UNMDB with GPU acceleration and NVMe/GPU‑direct tools demonstrate external‑memory pipelines for hundreds of TB.
AI-written summary · about
chapters (12)
- 0:00Scalability in data science
- 2:16Large dataset challenges
- 6:12NumPy and BLAS acceleration
- 10:27Exotic hardware exploration
- 13:23CPU vs GPU backends
- 15:51Performance benchmarking
- 20:50ML specialized libraries
- 23:20Tabular data benchmarking
- 26:30Python optimizations
- 29:19Hybrid systems future
- 32:48Data operations standards
- 34:16Q&A
related talks
31mcuDF: RAPIDS GPU-Accelerated Dataframe Library
02h 28mGPU Accelerated Python
01h 12mGPU-Accelerated Data Analytics in Python
02h 55mGPU Accelerated Python w: CUDA
30mCuPy: A NumPy-compatible Library for GPU
39mAccelerating Data Science with RAPIDS
25mScale Smarter, Not Harder, with cuPyNumeric.
29mcudf.pandas: The Zero Code Change GPU Accelerator for Pandas