Efficient and Portable 3D Explorable World Generation on AMD GPUs

2026-06-18 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

AMD has successfully optimized the open-source Matrix3D framework for explorable 3D world generation, enabling efficient execution on AMD Instinct™ MI250 and MI300 GPUs. Matrix3D, which combines panoramic generation with explicit 3D reconstruction for high-quality, coherent environments, saw its end-to-end generation time significantly reduced. On a single MI250 GPU, the time decreased from 2887s to 1306s, representing a 54% speedup. For the MI300 GPU, generation time dropped from 972s to 482s, a 50% improvement. These optimizations involved replacing CUDA-specific rendering kernels with portable Triton kernels, accelerating 3DGS fitting using the gsplat library, and refactoring the pipeline to reduce overhead from repeated model loading, I/O, and recomputation, alongside more efficient geometry optimization solvers.

Key takeaway

For AI engineers developing 3D content generation on AMD hardware, you should adopt the optimized Matrix3D pipeline to achieve significant performance gains. Your projects can benefit from the 50-54% latency reduction demonstrated on Instinct™ MI250 and MI300 GPUs. Consider integrating Triton kernels and gsplat for 3D Gaussian Splatting to enhance portability and efficiency, making your explorable world generation workflows faster and more accessible on ROCm-based systems.

Key insights

Optimizing 3D world generation on AMD GPUs requires kernel portability and pipeline efficiency.

Principles

• Explicit 3D representations yield better geometric consistency.
• Panoramic formulation offers broader spatial coverage.
• Cross-device portability improves framework accessibility.

Method

Replace CUDA kernels with Triton, use gsplat for 3DGS fitting, and refactor pipelines to minimize I/O and model loading overhead.

In practice

• Use Triton for portable, high-performance rendering kernels.
• Integrate gsplat for faster 3DGS reconstruction.
• Employ FFT/CG solvers for efficient depth map merging.

Topics

• 3D World Generation
• AMD GPUs
• ROCm
• 3D Gaussian Splatting
• Triton Kernels
• Performance Optimization

Code references

• ROCm/gsplat
• AMD-AGI/GEAK-agent
• microsoft/MoGe
• NVlabs/nvdiffrast
• facebookresearch/pytorch3d

Best for: Machine Learning Engineer, AI Engineer, AI Hardware Engineer

Related on AIssential

• Taking on CUDA With ROCm: ‘One Step After Another’ — AMD's ROCm software stack is rapidly maturing, leveraging open-source tools and community engagement to challenge Nvidia's CUDA dominance.
• Local-GS: Accelerating 3D Gaussian Splatting via Tile-Local Warp Coherence — Local-GS enhances 3DGS rendering by aligning Gaussian processing with GPU warp execution for improved efficiency.
• Unlocking Sparse Acceleration on AMD GPUs with hipSPARSELt — hipSPARSELt accelerates sparse matrix operations on AMD GPUs using 2:4 structured sparsity for improved AI model efficiency.
• Accelerating ComfyUI Workflows on AMD Instinct™ MI355X GPUs with ROCm — AMD Instinct MI355X GPUs significantly accelerate ComfyUI generative AI workflows over NVIDIA B200.
• World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible — World Tracing unifies pixel-aligned depth and generative completion for comprehensive 3D geometry reconstruction.
• Microsoft Research's Mirage gives video generation a persistent spatial memory that doesn't forget what's around the corner — Mirage uses latent spatial memory to achieve persistent spatial consistency and efficiency in video world models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.