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

DevFeed: [Mitigating SM-Level Load Imbalance in 3D Gaussian Splatting on GPUs](<https://devfeed.tech/articles/mitigating-sm-level-load-imbalance-in-3d-gaussian-splatting-on-gpus-125543.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.12319>)

Author: Junha Ryu, Hoi-Jun Yoo

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [gaussian splatting](<https://devfeed.tech/topics/gaussian-splatting.md>), [GPU optimization](<https://devfeed.tech/topics/gpu-optimization.md>), [scheduling](<https://devfeed.tech/topics/scheduling.md>), [3D rendering](<https://devfeed.tech/topics/3d-rendering.md>), [real-time 3D](<https://devfeed.tech/topics/real-time-3d.md>), [High-Performance Computing](<https://devfeed.tech/topics/high-performance-computing.md>)

Tags: [3d-gaussian-splatting](<https://devfeed.tech/tags/3d-gaussian-splatting.md>), [3dgs](<https://devfeed.tech/tags/3dgs.md>), [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [blending](<https://devfeed.tech/tags/blending.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-gr](<https://devfeed.tech/tags/cs-gr.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [gpu-scheduling](<https://devfeed.tech/tags/gpu-scheduling.md>), [gpus](<https://devfeed.tech/tags/gpus.md>), [kernel](<https://devfeed.tech/tags/kernel.md>), [lightweight](<https://devfeed.tech/tags/lightweight.md>), [load](<https://devfeed.tech/tags/load.md>), [multiple](<https://devfeed.tech/tags/multiple.md>), [outlier](<https://devfeed.tech/tags/outlier.md>), [real-time](<https://devfeed.tech/tags/real-time.md>), [rendering](<https://devfeed.tech/tags/rendering.md>), [schedule](<https://devfeed.tech/tags/schedule.md>), [sm](<https://devfeed.tech/tags/sm.md>), [statistics](<https://devfeed.tech/tags/statistics.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

## AI overview

Uneven work across screen tiles creates a long execution tail in GPU-based 3D Gaussian Splatting: a few SMs process heavy tiles while most sit idle, especially in large urban scenes. The proposed software scheduling approach uses a persistent kernel to prioritize heavy tiles and selectively splits outlier tiles across multiple CTAs. A per-frame policy uses tile statistics already available in the rendering pipeline to decide when and how much to split. The method requires neither retraining nor specialized hardware, preserves the original alpha blending formulation, and is described as compatible with existing optimizations across 3DGS variants. The abstract reports up to 2.93x average per-scene acceleration for alpha blending on a large urban scene and up to 5.6x on its most imbalanced frames.

## Source excerpt

arXiv:2611.12319v1 Announce Type: cross Abstract: 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.