NVIDIA/cutile-python

cuTile is a programming model for writing parallel kernels for NVIDIA GPUs

What it solves

cuTile Python provides a way to write high-performance GPU kernels using Python, allowing developers to express parallel computations on NVIDIA GPUs without needing to write low-level CUDA C++ code.

How it works

It acts as a programming language for NVIDIA GPUs that generates kernels based on Tile IR (Intermediate Representation). It allows users to define kernels using a @ct.kernel decorator and manage data movement between memory levels using load and store operations on tiles of data.

Who it’s for

Developers who want to program NVIDIA GPUs for high-performance computing and AI workloads, specifically those using Blackwell, Ampere, or Ada GPU architectures.

Highlights

  • Python-based GPU Programming: Write GPU kernels directly in Python.
  • Tile-based Memory Management: Explicit control over loading and storing data tiles for optimized performance.
  • Tile IR Integration: Leverages Tile IR to generate efficient GPU kernels.
  • Experimental Autotuning: Includes an experimental package for autotuning kernel launches.

Related

  • Project
  • Project
  • Project
  • Project
  • Project