GPU development with Python 101
summary
Claim: Python can run real CUDA without C++; tools used are numba CUDA kernels and cuDF/pandas plugin on NVIDIA’s DLI/Jupyter environment. Concrete results: a 30 million-element demo goes from ~8 s in pure Python to 30 million → 173 ms → 36 μs when memory stays on the GPU. Built kernels: elementwise double, a multi-pass image blur and a grayscale kernel (they work, but require bounds guards to avoid out‑of‑bounds artifacts). Takeaway: copy once, compute a lot, reuse device memory.
AI-written summary · about
chapters (5)
- 0:00GPU development in Python
- 15:26CPU versus GPU fundamentals
- 23:39First GPU kernel with Numba CUDA
- 55:20Image processing with GPU kernels
- 1:22:47Next steps and resources
related talks
02h 55mGPU Accelerated Python w: CUDA
48mGPU Programming in Pure Python
35mPython at the Speed of Light: A Complete Tour of the CUDA Ecosystem
33mWhen GPUs Make Python Slower: Patterns and Pitfalls
29mA Beginners Guide to GPUs for Pythonistas
27mSpeed up Python with hardware accelerators
34mWhat Python Developers Need to Know About Hardware: A Practical Guide...
55mAccelerated Python: The Community and Ecosystem