跳转至

从高斯点到硅片：重新审视 3DGS 的计算效率

文章背景与核心概要

3D 高斯泼溅（3D Gaussian Splatting, 3DGS）作为一种使用显式图元实现实时新视角合成的技术，近年来受到了学术界和工业界的广泛关注。然而，其系统级性能在不同的场景、视点、渲染路径以及硬件平台上往往表现出极大的差异。当前的许多研究试图通过改进表示方法、优化算法、微调 GPU 运行时或设计专用硬件来提升效率，但这些努力通常只针对渲染和更新流水线中的孤立节点，缺乏全局视角。

为了克服这一痛点，本文引入了一个以工作负载为中心（workload-centric）的框架，将算法设计、GPU 运行时以及硬件架构有机统一，从而识别出重复出现的计算模式。通过将文献分析与受控的 GPU 性能剖析（profiling）及复现测量相结合，作者证明了真正的系统级性能提升取决于三个关键因素：工作负载的减少是否能成功传导至下游执行、流水线各阶段之间是否存在适当的粒度匹配，以及如何有效管理数据传输、同步、缓存结果、梯度及优化器数据的开销。基于这些深刻的洞察，该研究指出了在渲染质量约束下实现更一致评估的潜在路径，并为未来的硬件和系统设计指明了关键方向。

目录

作者信息

Authors

  • Minnan Pei
  • Qiwei Dong
  • Yihan Zhou
  • Gang Li
  • Yuchen Zhu
  • Wenju Zhao
  • Zhongtian Long
  • Siting Wang
  • Peisong Wang
  • Jian Cheng

执行摘要

Executive Summary

3D Gaussian Splatting (3DGS) enables real-time novel-view synthesis using explicit primitives, yet its system-level performance varies dramatically across different scenes, viewpoints, rendering paths, and hardware platforms. Existing research attempts to optimize efficiency through representations, algorithms, GPU runtimes, and specialized hardware architectures; however, these efforts often target isolated points along the rendering and update pipelines.

This paper introduces a workload-centric framework that unifies algorithm design, GPU runtimes, and hardware architectures to identify recurring computational patterns. By combining literature analysis with controlled GPU profiling and reproduced measurements, the authors demonstrate that true system-level gains depend on: 1. Workload reductions propagating successfully to downstream execution. 2. Proper granularity matching across pipeline stages. 3. Managing the overhead of data transfers, synchronization, and cached results, gradients, and optimizer data.

Building on these insights, the study outlines paths toward more consistent evaluation under rendering-quality constraints and highlights vital directions for future hardware and system design.

摘要

Abstract

3D Gaussian splatting (3DGS) represents scenes with explicit primitives and supports real-time novel-view synthesis, yet its system efficiency varies substantially across scenes, viewpoints, rendering paths, and platform constraints. Existing studies pursue efficiency through representation and algorithm design, GPU runtime optimization, and architectural support, but their reported gains correspond to different points along the rendering and update paths. Connecting these indicators to end-to-end system benefit requires tracing how each optimization changes Gaussian selection, screen-space work, data movement, and stage or frame time. We therefore use a workload-centric framework to connect representation and algorithm research, GPU runtimes, and hardware architectures and to identify recurring workload patterns. We complement literature analysis with reproduced measurements and controlled GPU profiling of selected implementations, relating workload counts to stage time and memory traffic. Together, these comparisons show that system gains depend on workload reductions reaching downstream execution, granularity matching each stage, and the cost of data transfers, synchronization, and cached results, gradients, and optimizer data. Building on these findings, we discuss more consistent evaluation under rendering-quality constraints and identify key directions for future system design.

全文与参考链接

Full-Text & Reference Links