⚡ 本页包含 AI 生成的分析内容，仅供参考
本文提出IRIS，一款用于空间计算的SoC，支持基于3D高斯溅射的交互式逼真渲染和表面感知建模，解决了NeRF加速器无法支持表面提取和变形的问题，实现了8.55mJ/frame的低功耗性能。
applications, which demand a real-time and user-interactive 3D graphics system [1-3]. This requires real-time surface-aware modeling (SAM) to transfer a physical object to the virtual world, and interactive photorealistic rendering (IPR) is required to present the deformed virtual 3D object, as shown in Fig. 2.7.1 [4-5]. Neural Radiance Field (NeRF) accelerators [7-8] are unable to realize ISC since they cannot support surface extraction and deformation [9]. 3D Gaussian Splatting (3DGS) with free-deforming algorithms [10-11] enables high-quality rendering for deformed objects, but its realization on a high-power-consumption (>5W) edge GPU [12] requires >20minutes for SAM and >400ms for IPR due to its massive external memory access (EMA) and huge computational requirements. In general, there are
Seokchan Song, Seryeong Kim, Wonhoon Park, Jongjun Park, Sanghyuk An,
Gwangtae Park, Minseo Kim, Hoi-Jun Yoo KAIST, Daejeon, Korea Interactive Spatial Computing (ISC) enables immersive mixed reality (MR)