Research

AMD Optimizes Matrix3D for 3D World Generation on GPUs

AMD improved Matrix3D, a 3D world generation framework, with optimizations that cut end-to-end generation time by 54% on the MI250 GPU.

Image: AMD

AMD announced optimizations for Matrix3D, an explorable 3D world generation framework, to enhance efficiency and portability across AMD GPUs. The company deployed Matrix3D on Instinct™ MI250 and MI300 GPUs, achieving significant performance improvements. The end-to-end generation time for a single world was reduced from 2887 seconds to 1306 seconds on the MI250 GPU and from 972 seconds to 482 seconds on the MI300 GPU. These optimizations include replacing CUDA-specific rendering kernels with more portable Triton kernels, accelerating mesh optimization with more efficient solvers, and refactoring the pipeline to reduce overhead from repeated model loading and unnecessary I/O. According to AMD, these changes contribute to the overall speedup of the full pipeline, making the framework more accessible for developers and researchers working with AMD GPUs. Source: amd

AMD's optimizations for Matrix3D focus on improving both efficiency and portability. The company replaced the original rendering kernels with Triton kernels, which are more compatible across different hardware platforms. This change led to a 36% speedup over the nvdiffrast implementation while maintaining numerical accuracy. Additionally, AMD improved the 3DGS fitting process by replacing the original rasterization backend with the gsplat library, reducing the 3DGS fitting cost by 66%. These optimizations significantly contributed to the overall performance gains observed in the end-to-end generation process. Source: amd

The blog post provides detailed instructions for running Matrix3D on AMD GPUs, including system requirements and installation steps. Developers can use a pre-built Docker image and follow a script to clone the project and install the necessary dependencies. For text prompts, the script runs_m3d_t2i.sh is used, while for image prompts, the script runs_m3d_i2i.sh is employed. The post also includes qualitative examples of image-to-image and text-to-image generation results, illustrating the improvements in latency and performance. Source: amd

Key points

AMD improved Matrix3D, a 3D world generation framework, with optimizations that cut end-to-end generation time by 54% on the MI250 GPU.
AMD replaced CUDA-specific rendering kernels with more portable Triton kernels, achieving a 36% speedup over the nvdiffrast implementation.
AMD improved the 3DGS fitting process by replacing the original rasterization backend with the gsplat library, reducing the 3DGS fitting cost by 66%.
The end-to-end generation time for a single world was reduced from 2887 seconds to 1306 seconds on the MI250 GPU.
AMD's optimizations for Matrix3D include refactoring the pipeline to reduce overhead from repeated model loading and unnecessary I/O.

Source: AMD Read the original →

WRITTEN BY

Maya Chen

AI Research & Breakthroughs

Maya breaks down the latest AI research papers, benchmarks, and technical breakthroughs into plain language.