A 0.24mJ/Frame Quadratic Interpolation 4DGS Processor with Recursive Computation Reuse and Tree-Based Parallel-Rendering

📋 论文概要

该论文提出了一种用于4D高斯溅射(4DGS)的处理器，通过自适应二次插值、递归计算重用和树状并行渲染技术，解决了4DGS在内存需求、冗余计算和PE利用率低的问题。在D-Nerf数据集上实现了0.24mJ/帧的渲染能量，能效提升6.11倍。

💡 主要创新点

• 提出了自适应二次插值方法，减少内存需求并提高插值精度。
• 引入了递归计算重用技术，避免重复计算，降低计算冗余。
• 设计了基于树状结构的并行渲染架构，提升PE利用率，实现高效渲染。

🏷 关键词

📄 原文摘要

Abstract 4D Gaussian Splatting (4DGS) has widespread applications in fields such as VR, AR and industrial simulation. However, 4DGS suffers from significant memory requirements, redundant computations and low PE utilization. This paper introduces adaptive quadratic interpolation, recursive computation reuse and tree-based parallel rendering to tackle these challenges. The processor in the paper achieves a rendering energy of 0.24mJ/frame and improves energy efficiency by 6.11× on the D-Nerf dataset. 4D Gaussian Splatting (4DGS) is a pivotal technology for rendering dynamic scenes, with widespread applications in fields such as VR, AR and industrial simulation [1-5]. Unlike its static counterpart, 3D Gaussian Splatting (3DGS) [6-8], the attributes of 4DGS, like position, shape, and color, are functions of time, enabling photorealistic dynamic renderings [9-11]. The 4DGS workflow consists of two stages: pre-processing and rendering. In the preprocessing stage, a large number of 4DGS para

👥 作者与机构

Xiaolong Yang, Yang Wang, Wende Xu, Yubin Qin, Huanyu Wang, Ruiqi Guo, Zhiheng Yue, Jiangyuan Gu, Shaojun Wei, Yang Hu, Shouyi Yin

Tsinghua University, Beijing, China