Writing GPU kernels usually means leaving Python for CUDA C++ and losing the gradient on the way out. Warp keeps you in Python: decorate a function with @wp.kernel and it JIT-compiles to native CPU or GPU code that stays differentiable end-to-end, so the same simulation you run forward can also backpropagate.
What Sets It Apart
- Differentiability is first-class, not bolted on. Kernel gradients flow into PyTorch, JAX, and Paddle, so a physics step becomes just another differentiable layer in a training loop — directly useful for sim-to-real and gradient-based control.
- Simulation primitives ship in the box. Particles, meshes, ray-casting, FFT, and finite-element methods are built in, so you don't assemble a scientific-computing stack before writing the actual model.
- One source, two targets. The same kernel runs on CPU or GPU without a rewrite, and USD output drops results straight into standard visualization pipelines.
Who It's For
Great fit if you work on robotics, physics-based ML, or differentiable simulation and want gradients through GPU code without dropping to CUDA. Look elsewhere if you expect a high-level physics engine with ready-made rigid-body scenes — Warp is a kernel-authoring layer, so you build the simulation logic yourself, and reaching peak throughput still rewards understanding how the GPU executes your code.
