Mobile GPUs Get a 3D Boost with Texture-Optimized Gaussian Splatting

In a significant leap for mobile computing, researchers have developed Texture3dgs, a system that optimizes 3D Gaussian Splatting (3DGS) for mobile GPUs, enabling faster and more efficient 3D scene reconstruction directly on devices. This breakthrough addresses the growing demand for real-time applications like augmented reality and robotics, which require low-latency processing without relying on cloud servers. By focusing on mobile hardware constraints, such as limited memory bandwidth and specialized texture caches, the team achieves up to 1.7× end-to-end speedup and 1.6× memory reduction, pushing the boundaries of on-device AI and graphics. are vast, from enhancing user privacy by keeping data local to enabling responsive, offline-capable systems in dynamic environments.

Ology centers on a novel sorting algorithm tailored for mobile GPUs' 2D texture memory, which dominates computations in 3DGS pipelines. Traditional sorting s, like GPUTeraSort, struggle with cache inefficiencies when handling large datasets, leading to performance bottlenecks. Texture3dgs introduces a layout transformation that ensures comparing pairs in sorting steps are adjacent in memory, reducing cache misses by up to 60% in L1 cache. Additionally, the system employs variable packing to organize Gaussian parameters—such as position, opacity, and color coefficients—into efficient textures, minimizing data movement. Other optimizations include stage fusion to combine kernel operations and tile-based rendering enhancements that leverage SIMD execution, all validated through a cost model based on empirical profiling of texture cache behavior.

From extensive evaluations on off-the-shelf mobile platforms, including Snapdragon 8 Gen 2, demonstrate that Texture3dgs outperforms state-of-the-art baselines like 3dgs.cpp and TensorFlow Lite. The optimized sorting algorithm alone achieves up to 4.1× speedup over GPUTeraSort, with latency reductions across various scene complexities from datasets like Tanks and Temples and Synthetic-NeRF. In end-to-end tests, Texture3dgs reduced overall latency by an average of 1.25×, with peaks of 1.7× in less complex scenes, while cutting memory access by 25% and peak usage by 20%. Portability tests on older devices like Xiaomi MI 6 and Redmi Note 10 confirmed stable performance, underscoring the robustness of the approach across different mobile GPU architectures.

Of this research extend to numerous fields, empowering real-time 3D applications in augmented reality for immersive experiences, robotics for obstacle avoidance, and autonomous systems requiring rapid environmental modeling. By enabling efficient on-device processing, Texture3dgs supports data privacy and offline functionality, critical for applications in sensitive or remote settings. It also sets a precedent for hardware-aware algorithm design, encouraging further innovations in mobile AI that could revolutionize industries from gaming to industrial automation. The study highlights how optimizing for specific memory hierarchies can unlock performance gains without hardware upgrades, making advanced graphics accessible on everyday devices.

Despite its successes, the work has limitations, such as reliance on floating-point sorting that required key normalization for 64-bit integers in 3DGS, potentially introducing precision trade-offs. The optimizations are tailored for mobile GPUs with 2D texture caches and may not directly translate to other architectures like NPUs or desktop systems. Future work could explore adaptive resolution models, integration with compression techniques like those in LightGaussian, and extensions to emerging mobile processors to broaden applicability. Overall, Texture3dgs represents a pivotal step toward practical, real-time 3D reconstruction on resource-constrained devices, with code and evaluations paving the way for community adoption and refinement.