When GPUs Make Python Slower: Patterns and Pitfalls
summary
GPU acceleration often slows code when fixed costs dominate: moving RAM↔VRAM wipes out gains so regex string search on cuDF is slower for small or high-bandwidth jobs, while dense math like matrix multiplication on GPU is much faster (compute‑bound). Converting rowwise Python loops creates millions of tiny kernels and yields ~1–3% GPU utilization, surprising 3× vs expected 300× speedups. Tools shown: Numba, CuPy, cuDF, cuGraph and Nsight profiling; rules: minimize memory transfers, maximize operational intensity, avoid many small kernels.
AI-written summary · about
chapters (7)
- 0:00Introduction and goals
- 1:23Minimizing memory transfers
- 4:28Analyzing GPU suitability
- 12:38Operational intensity
- 18:43Data format impact
- 24:03Avoiding kernel overhead
- 28:12Conclusion and Q&A
related talks
02h 28mGPU Accelerated Python
02h 55mGPU Accelerated Python w: CUDA
29mA Beginners Guide to GPUs for Pythonistas
30mCUDA in Your Python: Effective Parallel Programming on the GPU
01h 02mExploit your GPU power with PyCUDA (and friends)
34mAccelerating GPU Algorithms in pure Python
01h 29mGPU development with Python 101
02h 14mGPU development in Python 101