CUDA in Your Python: Effective Parallel Programming on the GPU
summary
Shows a concrete vector-add benchmark that yields 30× speedup by swapping NumPy for a GPU-backed drop-in, and names the actual tools: CuPy drop-in replacement, Numba CUDA JIT, plus PyCUDA/PyCUDA-style raw kernels, RAPIDS, nvcc and Cython extensions. Explains what breaks: strings and Python objects aren’t supported, CUDA errors can appear asynchronously, and grid/block tuning plus explicit host/device copies and shared memory govern performance. Mentions building a two‑GPU home server after an unexpected cloud bill.
AI-written summary · about
chapters (22)
- 0:00Start
- 1:02Moore's Law Is Dead
- 3:21History of the Gpu
- 3:48Specs
- 5:35Cpu versus Gpu
- 8:11Horton's Law
- 8:37Gpus for Deep Learning
- 9:52How Can You Start Programming the Gpu
- 11:12Data Types
- 12:22Threads Blocks and Grids
- 15:44Host and Device
- 18:36Cuda Simulator
- 18:59Interoperability
- 19:28Pi Cuda
- 20:25Memory Management
- 21:39Meta Programming
- 22:05Templating Library
- 23:24Extending Python
- 23:43Main Challenges
- 25:21How Do You Connect Your Cuda Code
- 27:04Manual Memory Management
- 27:59Accessing a Gpu
related talks
01h 02mExploit your GPU power with PyCUDA (and friends)
01h 02mExploit your GPU power with PyCUDA (and friends)
02h 28mGPU Accelerated Python
35mPython at the Speed of Light: A Complete Tour of the CUDA Ecosystem
35mCUDA in Python- A New Era for GPU Acceleration
34mAccelerating GPU Algorithms in pure Python
48mGPU Programming in Pure Python
33mWhen GPUs Make Python Slower: Patterns and Pitfalls