Python at the Speed of Light: A Complete Tour of the CUDA Ecosystem
summary
Shows how to accelerate Python numerical code with the CUDA ecosystem: migrating NumPy code to CuPy, using RAPIDS dataframes (Pandas/Polars) for GPU-accelerated dataframes, and writing GPU kernels with Numba CUDA and CUDA Python 1.0 + MLIR. Demonstrates concrete examples—vector add, matrix multiply, and a convolution that calls FFT inside a kernel that compiles and runs—and explains CCCL, host vs device APIs, CUDA streams, and cuTile as a NumPy-like interface to tensor cores, arguing for kernel fusion based on a 30–60 flops-per-byte heuristic.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:52What is CUDA Python
- 3:41Basic CUDA kernel
- 5:50Manual to optimized kernels
- 7:37CUDA Python layers
- 9:48Matrix multiplication Numba
- 11:53Block programming CuPy
- 13:17CCCL GPU algorithms
- 15:25Linear algebra libraries
- 17:50Kernel fusion memory
- 19:14Multi-GPU communication
- 20:42Key libraries summary
- 21:24Advanced CUDA Core
- 22:46Why Python for CUDA
- 23:29Domain-specific libraries
- 25:14Rapids Python integration
- 26:17Beyond PyTorch TensorFlow
- 26:59Q&A
related talks
30mCUDA in Your Python: Effective Parallel Programming on the GPU
02h 28mGPU Accelerated Python
55mAccelerated Python: The Community and Ecosystem
02h 55mGPU Accelerated Python w: CUDA
01h 29mGPU development with Python 101
35mCUDA in Python- A New Era for GPU Acceleration
34mAccelerating GPU Algorithms in pure Python
01h 02mExploit your GPU power with PyCUDA (and friends)