NVIDIA Warp

NVIDIA's Python framework that compiles ordinary Python functions into GPU code for writing simulations.

Warp is an open-source Python framework from NVIDIA for writing high-performance simulation and graphics code. Developers write ordinary Python functions and mark them with the @wp.kernel decorator; Warp then just-in-time compiles them into CUDA or CPU code that runs in parallel, so there is no need to write C++/CUDA by hand. It also supports automatic differentiation, so a simulation written in Warp can be differentiated with respect to its inputs — an approach known as differentiable simulation. Warp includes built-in tools for geometry, collision, and particles, and can exchange data directly with PyTorch and JAX. NVIDIA's Newton physics engine and MuJoCo Warp are both built on top of it.

ExampleWrite a cloth simulation in a few dozen lines of Warp code, advance thousands of parallel environments at once on the GPU, and backpropagate gradients into PyTorch to optimize a grasping trajectory.

Also called

Warp

Related

Newton Physics Engine · MuJoCo Warp · Differentiable Simulation · CUDA · GPU-Accelerated Parallel Simulation · Operator / Kernel

Sources

NVIDIA/warp (GitHub)

As of

2025