Video World Models · 2026-11-10

Mitigating SM-Level Load Imbalance in 3D Gaussian Splatting on GPUs

Junha Ryu, Hoi-Jun Yoo

arXiv:2611.12319PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors identify SM-level load imbalance in GPU-based 3D Gaussian Splatting (3DGS) alpha blending caused by uneven per-tile workloads, which creates a long execution tail. They propose software scheduling: persistent-kernel LPT dispatch of heavy tiles and selective depth-wise chunking with saturation signaling. A lightweight per-frame policy decides chunking. They report alpha blending speedups up to 2.93× averaged on a large-scale urban scene and up to 5.6× on the most imbalanced frames.

Why it matters

GPU utilization issues in 3DGS alpha blending can limit real-time rendering in settings like large-scale urban scenes and driving simulators. The method targets execution-tail reduction at the rendering system level without retraining or specialized hardware, and is stated to apply orthogonally to multiple 3DGS variants and optimized rasterizers, while retaining the original alpha blending formulation.

Method

  • Use a persistent kernel to dispatch tiles in longest-processing-time-first (LPT) order based on per-tile projected Gaussian counts, so heavy tiles start earlier and reduce the execution tail.
  • For outlier tiles that remain bottlenecks, partition their depth-ordered Gaussian list into depth-contiguous chunks processed in parallel across multiple CTAs, combined with a reduction step.
  • Selectively chunk only the heaviest tiles and co-execute the chunk kernel with the LPT persistent kernel on separate CUDA streams; use a lightweight per-frame policy to choose how many tiles and how deeply to chunk using tile statistics already available in the pipeline.

Limitation

The authors state that their approach targets the alpha blending stage of inference-time rendering, and its benefit scales with the tile imbalance of each frame; on balanced workloads the execution tail is short and there is limited room for improvement.

Abstract (from arXiv)

3D Gaussian Splatting (3DGS) enables photorealistic, real-time rendering on commodity GPUs and is increasingly adopted in generative world models and autonomous driving simulators. However, existing optimizations rely on algorithmic changes that often require retraining, or on specialized accelerators that lack flexibility across evolving 3DGS variants. In this paper, we identify a fundamental inefficiency in GPU-based 3DGS rendering: severe SM-level load imbalance caused by uneven workloads across screen tiles. This imbalance creates a long execution tail, leaving most SMs idle while a few process computationally intensive tiles, particularly in large-scale urban scenes. To address this, we present a software-defined GPU scheduling approach that mitigates the execution tail. Our approach uses a persistent kernel to schedule heavy tiles first and selectively chunks outlier tiles across multiple CTAs. A lightweight per-frame policy determines when and how much to chunk using tile statistics already available in the rendering pipeline. Our approach requires neither retraining nor specialized hardware and can be applied orthogonally to existing optimizations across diverse 3DGS variants. Averaged per scene, it accelerates alpha blending by up to 2.93x on a large-scale urban scene, and by up to 5.6x on its most imbalanced frames, while retaining the original alpha blending formulation.

Related papers