Bringing NVIDIA math libraries to Python scientific community
summary
nvmath-python packages NVIDIA math libraries into Python with JIT fusion so user-defined python device callbacks compile via Numba or Warp and run as single kernels, claiming 90% peak performance on hardware. It exposes zero-copy array interop via DLP, provides host and device APIs, adds CPU execution space and RNG, and ships on conda/pip and GitHub. A recent proof of concept demonstrates multi-gpu multi-node scale to thousands of gpus for FFTs and the project reports incremental releases since its July debut.
AI-written summary · about
chapters (9)
- 0:00NVIDIA math libraries in Python
- 1:28Need for new Python math library
- 3:38Pythonic JIT for CUDA
- 7:14Core design of nvMath Python
- 12:20Python ecosystem interop
- 14:54nvMath Python architecture
- 18:50Roadmap and releases
- 22:05Summary and future vision
- 25:01Q&A
related talks
35mCUDA in Python- A New Era for GPU Acceleration
34mAccelerating GPU Algorithms in pure Python
55mAccelerated Python: The Community and Ecosystem
30mNumba, a JIT compiler for fast numerical code
02h 28mGPU Accelerated Python
30mCUDA in Your Python: Effective Parallel Programming on the GPU
27mSpeed up Python with hardware accelerators
37mUsing Numba Effectively Today