---
格式版本: 2
标题: "Atlas: Algorithm-Hardware Co-Design for On-Device City-Scale 3D Gaussian Splatting in VR"
原文链接: "https://arxiv.org/abs/2609.02352"
发布日期: "2026-09-02"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Wed, 2 Sep 2026 09:25:00 UTC (6,266 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-04T01:34:21+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-04T01:32:14+08:00"
入库时间: "2026-09-03T17:34:21.473Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=GPU&searchtype=all"
匹配关键词:
  - "GPU"
  - "performance"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 19
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该资料为arXiv论文，主题为VR端城市规模3D高斯泼溅渲染，属于移动GPU与算法设计，与超节点/AI Rack/数据中心基础设施完全无关，无任何相关技术或商业信号。"
AI质检模型: "zj-deepseek-v4-flash"
AI质检时间: "2026-09-04T01:34:38+08:00"
AI主题相关性: 0
AI来源权威性: 5
AI新颖性: 3
AI技术细节: 3
AI商业部署信号: 0
AI完整性: 8
AI摘要: "Atlas提出一种面向移动VR设备端城市规模3D高斯泼溅渲染的算法-硬件协同设计框架，利用每帧仅需当前视角和细节层级下部分高斯数据的特点，实现分层内存动态卸载，无需运行时联网即可扩展渲染规模。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-04T00:18:10.011Z"
采集批次: "2026年9月3日22点43分34秒"
采集批次ID: "20260903-224334-406"
去重键: "https://arxiv.org/abs/2609.02352"
---

## Computer Science > Hardware Architecture

## Title:Atlas: Algorithm-Hardware Co-Design for On-Device City-Scale 3D Gaussian Splatting in VR

Authors:[He Zhu](https://arxiv.org/search/cs?searchtype=author&query=Zhu,+H), [Zheng Liu](https://arxiv.org/search/cs?searchtype=author&query=Liu,+Z), [Xingyang Li](https://arxiv.org/search/cs?searchtype=author&query=Li,+X), [Anbang Wu](https://arxiv.org/search/cs?searchtype=author&query=Wu,+A), [Zihan Liu](https://arxiv.org/search/cs?searchtype=author&query=Liu,+Z), [Ruyang Li](https://arxiv.org/search/cs?searchtype=author&query=Li,+R), [Hui Wei](https://arxiv.org/search/cs?searchtype=author&query=Wei,+H), [Yaqian Zhao](https://arxiv.org/search/cs?searchtype=author&query=Zhao,+Y), [Jingwen Leng](https://arxiv.org/search/cs?searchtype=author&query=Leng,+J), [Minyi Guo](https://arxiv.org/search/cs?searchtype=author&query=Guo,+M), [Yu Feng](https://arxiv.org/search/cs?searchtype=author&query=Feng,+Y)

[View PDF](https://arxiv.org/pdf/2609.02352) [HTML (experimental)](https://arxiv.org/html/2609.02352v1)

> Abstract:3D Gaussian splatting (3DGS) has drawn significant attention in the architectural community recently. However, enabling city scale 3DGS on mobile VR devices remains challenging, as the memory requirement of large scale scenes far exceeds the memory capacity of today's mobile GPUs. This paper presents Atlas, an on device city scale 3DGS rendering framework that enables scalable rendering without runtime Internet access. The key insight is that although the full 3DGS model is massive, each frame only requires a small subset of Gaussians under the current pose and level of detail requirement. Based on this insight, Atlas introduces a hierarchical memory offloading mechanism that dynamically loads only necessary Gaussian data into device memory. To further improve performance, Atlas proposes temporal aware LoD search and stereo rasterization to avoid redundant computation in VR. We further show that our technique can be integrated with existing 3DGS accelerators with negligible hardware overhead. Overall, Atlas achieves 18.5x speedup over the GPU baseline and 3.9x speedup over the state of the art 3DGS accelerators, with 92.4% energy savings.

| Subjects: | Hardware Architecture (cs.AR) |
| --- | --- |
| Cite as: | [arXiv:2609.02352](https://arxiv.org/abs/2609.02352) \[cs.AR\] |
|  | (or [arXiv:2609.02352v1](https://arxiv.org/abs/2609.02352v1) \[cs.AR\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2609.02352](https://doi.org/10.48550/arXiv.2609.02352) |

## Submission history

From: Yu Feng \[[view email](https://arxiv.org/show-email/2a692c73/2609.02352)\]  
**\[v1\]** Wed, 2 Sep 2026 09:25:00 UTC (6,266 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2609.02352) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))