NVIDIA Releases Warp v1.10 with Enhanced JAX Integration and Performance Improvements

NVIDIA has announced the release of Warp v1.10, introducing significant enhancements to JAX integration, performance optimizations, and new features aimed at AI and graphics developers. This update, released on November 2, 2025, expands the capabilities of Warp by integrating automatic differentiation support with JAX and improving multi-device compatibility.

The latest version also includes performance improvements in BVH operations, faster built-in function calls from Python, and usability enhancements such as negative indexing and slicing for arrays. Additionally, Warp v1.10 introduces new built-in functions, including error functions and type casting, to streamline development workflows.

Key Features of Warp v1.10

Enhanced JAX Integration

Warp v1.10 focuses on strengthening its integration with JAX, a popular library for machine learning research. The update introduces experimental support for automatic differentiation with JAX, enabling developers to compute gradients through Warp kernels using jax.grad(). This feature allows for more efficient optimization and training of AI models.

The release also includes proper support for jax.pmap() and jax.shard_map(), enabling multi-device parallel execution. This improvement allows developers to scale their workloads across multiple GPUs, accelerating computation and improving performance.

Performance and Usability Improvements

Performance enhancements are a key focus of Warp v1.10. The update introduces up to 70x faster built-in function calls from Python, significantly reducing execution time for critical operations. Sparse matrix and finite element method (FEM) operations can now be captured in CUDA graphs, further optimizing performance for complex simulations.

Usability improvements include support for negative indexing and improved slicing behavior for arrays, making array manipulation more intuitive. The release also introduces a new wp.Bvh.rebuild() method, allowing BVH hierarchies to be rebuilt in-place without allocating new memory, with full CUDA graph support.

Deprecations and Platform Support

Notably, Warp v1.10 removes the deprecated warp.sim module, which has been superseded by the Newton physics engine. Users relying on warp.sim are advised to migrate to Newton using the provided migration guide.

Support for Intel-based macOS (x86_64) has been discontinued, while Apple Silicon Macs (ARM64) remain fully supported. Additionally, the release plans to drop support for Python 3.8 in the next minor update, encouraging developers to upgrade to Python 3.9 or later.

Conclusion

Warp v1.10 represents a significant step forward for AI and graphics development, offering enhanced JAX integration, performance improvements, and usability features. With these updates, NVIDIA continues to empower developers to build more efficient and scalable applications.