Luxuries, necessities, and the challenges that remain
summary
Shows concrete results using Nvidia hardware and Python: cuPy and RAPIDS let UMAP run at interactive latency so visual exploration uses full data instead of tiny subsamples, XGBoost on GPU yields about a 10x speedup versus CPU, self‑organizing maps get ~3x–6x faster on GPU, and a workstation GeForce suffices for many tasks. Also reports a negative: a discrete‑event simulation stalls because missing SciPy RNG and distribution support forces a costly refactor for full GPU portability.
AI-written summary · about
chapters (12)
- 0:00Introduction and community
- 1:24What is data science
- 3:51History of acceleration
- 9:26When acceleration is necessary
- 13:14Acceleration libraries overview
- 15:00Case study visualization
- 19:56Case study model training
- 22:43Case study prototype acceleration
- 29:24Case study model explainability
- 32:12Case study negative result
- 37:48Scaling out and challenges
- 39:11Summary and takeaways
related talks
01h 12mGPU-Accelerated Data Analytics in Python
02h 55mGPU Accelerated Python w: CUDA
01h 08mGPU Data Science with Rapids
02h 28mGPU Accelerated Python
41mAccelerating Data Science with RAPIDS
27mThe art of wrangling your GPU Python environments
29mA Beginners Guide to GPUs for Pythonistas
01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUs