NVIDIA/warp

A Python framework for GPU-accelerated simulation, robotics, and machine learning.

What it solves

NVIDIA Warp is designed to bridge the gap between high-level Python productivity and the high-performance requirements of GPU-accelerated simulations. It allows developers to write simulation code in Python while achieving the execution speed of compiled kernel code on CPUs or GPUs.

How it works

Warp uses Just-In-Time (JIT) compilation to transform regular Python functions into efficient kernel code. It provides a specialized set of primitives for geometry processing, robotics, and physics simulation. Crucially, these kernels are differentiable, meaning they can be integrated directly into machine learning pipelines using frameworks like PyTorch, JAX, and Paddle.

Who it’s for

It is intended for researchers and engineers working in robotics, computational physics, and machine learning who need to build high-performance, differentiable simulations.

Highlights

  • JIT Compilation: Converts Python functions to GPU/CPU kernels.
  • Differentiable Kernels: Supports integration with ML frameworks for gradient-based optimization.
  • Rich Primitive Set: Includes built-in tools for physics, robotics, and geometry processing.
  • Cuda Acceleration: Optimized for NVIDIA GPUs to handle massive parallel simulations (e.g., million-particle systems).
  • Cross-Platform: Supports Windows, Linux, and macOS (CPU execution).

Related

  • Project
    newton-physics/newton

    Newton is a GPU-accelerated, differentiable physics simulation engine built on NVIDIA Warp and MuJoCo Warp, designed for scalable robotics research.

  • Project
    brainpy/BrainPy

    A JIT-compiled framework for computational neuroscience and brain-inspired computation that enables efficient simulation, training, and analysis of brain dynamics models.

  • Project
    facebookincubator/AITemplate

    A Python framework that transforms deep neural networks into optimized CUDA/HIP C++ code for high-performance inference serving on NVIDIA and AMD GPUs.

  • Project
    Jittor/jittor

    Jittor is a high‑performance deep‑learning framework that JIT‑compiles the entire computation graph and uses meta‑operators to generate specialised CUDA/C++ kernels. It offers a PyTorch‑like Python API, supports CPU and GPU back‑ends (CUDA, ROCm, Hygon), and includes a model zoo for vision, rendering, geometric learning, and reinforcement learning.

  • Project
    NVIDIA/cuEquivariance

    An NVIDIA Python library for building high-performance geometric neural networks that respect 3D symmetries using optimized CUDA kernels and segmented polynomials.