NVIDIA Warp
NVIDIA Warp is a purpose-built open source Python framework that delivers GPU acceleration for computational physics, AI, and optimization workflows.
How NVIDIA Warp Works
Supercharging Physics Simulation
NVIDIA Warp gives coders an easy way to write GPU-accelerated programs for simulation AI, robotics, and machine learning (ML). With NVIDIA Warp, Python developers can create accelerated, differentiable simulation workflows that drive ML pipelines in PyTorch, JAX, NVIDIA PhysicsNeMo, and NVIDIA Omniverse™. Benefits include simulation performance on par with native CUDA® code, with the convenience and developer productivity of Python.
Kernel-Based Code
NVIDIA Warp performs a just-in-time (JIT) runtime compilation of Python functions to CUDA kernel-level and x86 code. Kernel-based programming provides a low-level abstraction that maps closely to GPU hardware, and, in contrast to tensor-based programming, provides implicit kernel fusion (controlled by the user), fine-grained control over threads, native support for conditional logic, and sparse scatter and gather common in simulation code.
Differentiable Programming
In addition to generating forward-mode kernel code, Warp can generate reverse-mode (adjoint) kernels that propagate the gradients of simulation results back into frameworks, such as PyTorch and JAX for network training, design optimization, and parameter estimation.
Tile-Based Programming
NVIDIA Warp provides a block-based programming model where threads cooperate to perform operations on data tiles. This abstraction allows kernels to leverage dedicated hardware units, such as Tensor Cores, for high-performance matrix multiplication and Fourier transforms, while enabling developers to optimize data movement between global, shared, and register memory for accelerated scientific computing.
Native Geometry Primitives
NVIDIA Warp provides high-performance data structures essential for simulation and graphics. Developers can leverage triangle meshes, sparse volumes (NanoVDB), and spatial acceleration structures like hash grids and bounding volume hierarchies (BVHs). These primitives are optimized to accelerate complex geometric queries such as raycasts and nearest neighbor searches.
Sparse Linear Algebra
NVIDIA Warp supports sparse linear algebra operations essential for simulation. It provides efficient Block Sparse Row (BSR) and Compressed Sparse Row (CSR) matrix formats, along with preconditioned iterative solvers such as conjugate gradient (CG) and GMRES optimized for GPU execution.
Finite Element Method (FEM) Toolkit
NVIDIA Warp provides a dedicated module for solving differential equations using finite-element methods. It enables users to define integrals over domains, assemble sparse linear systems, and solve them using built-in solvers. The module supports various mesh types and high-order function spaces, allowing for rapid experimentation with custom formulations for diffusion, elasticity, and fluid flow.
Starter Kits
Many Python developers are using NVIDIA Warp today for physics simulation, solver development, data processing, and visualization. NVIDIA Warp includes several higher-level data structures and primitives that make implementing simulation and geometry processing algorithms easier.
Accelerate CAE Tool Development
Develop next‑generation CAE solvers that run interactively, combining NVIDIA Warp kernels with your existing simulation code to unlock GPU‑accelerated, AI‑ready digital twins and optimization loops.
Computational Physics
Prototype and scale custom solvers for rigid bodies, fluids, and elastic materials in Python, using NVIDIA Warp kernels that JIT‑compile to CUDA for production‑grade performance on GPUs.
Robotics Simulation
Run high-fidelity robot motion planning and control pipelines with GPU-accelerated, differentiable physics built on NVIDIA Warp, then deploy policies to real robots with confidence in sim‑to‑real performance.
Training and Optimization
Integrate differentiable simulations into ML workflows to optimize controllers, physical parameters, and designs end‑to‑end with gradient‑based methods using NVIDIA Warp and popular learning frameworks.
Geometry Processing
Build high‑performance geometry and mesh processing pipelines in Python, using NVIDIA Warp’s spatial computing primitives for tasks like meshing, remeshing, collision queries, and distance field operations.
Newton Physics Engine
Use Newton, an open source, GPU‑accelerated physics engine built on NVIDIA Warp and OpenUSD, to create extensible, differentiable simulation environments for robotics and reinforcement learning at scale.
NVIDIA Warp On Demand Playlist
Resources
Get started with NVIDIA Warp today.