ROCm-Simulation 25.11 Release notes

ROCm-Simulation 25.11 Release notes#

2 min read time

Applies to Linux

This is the second release of the AMD ROCm Simulation Domain toolkit (ROCm-Simulation), a comprehensive open-source software collection designed to accelerate physics-based and numerical simulations on AMD GPUs.

Release highlights#

This release introduces support for ROCm 7.0.0 for the two components:

  • Taichi Lang is an open-source, imperative, parallel programming language for high-performance numerical computation. It is embedded in Python and uses just-in-time (JIT) compiler frameworks (such as LLVM) to offload the compute-intensive Python code to the native GPU or CPU instructions.

  • GSplat (Gaussian splatting) is a highly efficient technique for real-time rendering of 3D scenes trained from a collection of multiview 2D images of the scene. It has emerged as an alternative to neural radiance fields (NeRFs), offering significant advantages in rendering speed while maintaining visual quality.

System requirements#

Note

For the 25.11 release, the ROCm-Simulation components both require ROCm 7.0.0. Ensure you have the system requirements listed in the installation instructions to proceed for each component. If you are only installing Taichi Lang or GSplat, see the compatibility matrix.

ROCm-Simulation components#

The following table lists ROCm-Simulation components versions for ROCm-Simulation 25.11, including any version changes for the components. Click to go to the component’s source code on GitHub.

Name Version
Taichi Lang 1.8.0b1 ⇒ 1.8.0b2
GSplat 1.5.3b1 ⇒ 1.5.3b2

Detailed component changelogs#

Taichi Lang 1.8.0b2#

This release adds support for ROCm 7.0.0.

Latest features:

  • Added support for AMD Instinct MI355X, MI325X, and MI300X GPUs.

GSplat 1.5.3b2#

This release adds support for ROCm 7.0.0.

Latest features:

  • Added support for AMD Instinct MI325X GPUs.

  • Fused SSIM is enabled: Fused SSIM refers to an optimized implementation of the Structural Similarity Index Measure (SSIM) that combines multiple operations into a single, more efficient kernel. Benchmarks show a resultant significant improvement in training time.