NVIDIA has announced the release of Warp v1.10 on November 2, 2025, which expands JAX integration with automatic differentiation support and multi-device jax.pmap() compatibility. The update also includes enhancements to the tile programming model, performance improvements in BVH operations, and faster built-in function calls from Python. Additional usability improvements feature negative indexing and slicing for arrays, as well as new built-in functions including error functions and type casting. Warp v1.10 focuses on improving performance, usability, and integration with other tools in the AI and graphics development ecosystem. Key updates include:

JAX Automatic Differentiation: Experimental support for automatic differentiation with JAX, enabling computation of gradients through Warp kernels using jax.grad().
Multi-device JAX Support: Proper support for jax.pmap() and jax.shard_map() for multi-device parallel execution.
In-place BVH Rebuilding: A new wp.Bvh.rebuild() method allows rebuilding BVH hierarchies in-place without allocating new memory, with CUDA graph support.
Tile Programming Enhancements: Axis-specific reductions, component-level indexing, and convenience functions for creating tiles.
Performance Improvements: Up to 70x faster built-in function calls from Python and sparse matrix/FEM operations that can be captured in CUDA graphs.
Array Indexing and Slicing: Support for negative indexing and improved slicing behavior for more intuitive array manipulation.

Notably, this release removes the `warp.sim` module, which was deprecated since v1.8 and is now superseded by the Newton physics engine. Users relying on `warp.sim` should migrate to Newton, consulting the provided migration guide for assistance. Support for Intel-based macOS (x86_64) has been removed, while Apple Silicon Macs (ARM64) continue to be fully supported. The release also plans to drop support for Python 3.8 in the next minor release.