CUDA in Python

What if your Python code could achieve over 90% of a GPU's theoretical max performance? Learn how NVIDIA is making it possible.

#1about 6 minutes

Understanding the CUDA platform stack for Python developers

The CUDA platform is layered from high-level domain libraries to low-level hardware access, with new tools aiming to combine Python's productivity with GPU performance.

#2about 3 minutes

Improving performance by fusing GPU operations

The nvmath-python library enables kernel fusion using epilogues, which combines multiple operations like matrix multiplication and bias addition into a single GPU kernel launch.

#3about 5 minutes

Calling device-side functions directly from Python kernels

Python kernels can now directly call pre-compiled, high-performance device-side functions from libraries like cuBLAS, enabled by a just-in-time linker called nvJitLink.

#4about 2 minutes

Fine-grained parallelism with cooperative groups in Python

The CUB library is exposed to Python, allowing for cooperative operations and reductions at the block or warp level for fine-grained control over GPU parallelism.

#5about 3 minutes

Accelerating language support with numba-cuda and nupack

The numba-cuda module is separated to accelerate feature delivery, while nupack automatically generates Python bindings for C++ templated code.

#6about 4 minutes

A Pythonic object model for host-side GPU control

A new high-level object model allows Python developers to directly manage GPU resources like devices, contexts, streams, and linker objects without boilerplate code.