NVIDIA Warp is a Python framework for GPU-accelerated simulation, robotics, and machine learning.
Warp solves the problem of writing high-performance GPU code without leaving Python by taking regular Python functions and JIT compiling them to efficient kernel code that runs on CPU or GPU. The framework provides a rich set of built-in primitives for physics simulation, robotics, and geometry processing, eliminating the need to hand-write low-level CUDA kernels for common tasks. Kernels are differentiable by design, allowing them to integrate directly into machine-learning pipelines with PyTorch, JAX, and Paddle.
Warp suits projects that combine simulation with learning or optimization, particularly in robotics, physics-based graphics, and scientific computing where GPU acceleration is essential. It works well for teams that want to stay in the Python ecosystem while achieving performance comparable to hand-written CUDA code. The framework supports Windows, Linux, and macOS, with CUDA acceleration available on Windows and Linux for NVIDIA GPUs, while all platforms support CPU execution. Installation is straightforward via PyPI, and the project provides tutorial notebooks covering introductory concepts through differentiable simulations.
The project maintains an active development cadence with regular updates to the changelog. The codebase shows consistent refinement of core functionality and expansion of supported use cases. Documentation is comprehensive and kept current alongside releases. The project demonstrates responsiveness to user needs through its tutorial offerings and example coverage across simulation domains.
