Lune

ISCA2026顶会

GauTracer: Extending Ray Tracing Accelerator for Gaussian-Based Scene Representation

Lizhou Wu, Kunchen Zou, Yuzheng Lin, Chixiao Chen, Xiaoyang Zeng, Haozhe Zhu

2026年份

摘要

Gaussian splatting has emerged as a promising paradigm in 3D vision, enabling instant scene reconstruction and photorealistic rendering. However, integrating rasterization-oriented Gaussian representations into ray tracing pipelines remains challenging. Existing ray tracing accelerators (RTAs) lack native support for Gaussian primitives, forcing reliance on software shader, which incurs significant instruction overhead, thread divergence, and excessive memory traffic. To address this, we present GauTracer, a ray tracing accelerator with efficient hardware support for Gaussian primitives. GauTracer elevates Gaussians to first-class primitives via lightweight microarchitectural extensions, including a dedicated Ray-Gauss Intersection Unit (RGIU). To improve versatility, the RGIU adopts a reconfigurable design supporting both 3D ellipsoidal and 2D planar Gaussian primitives with minimal overhead. In addition, a max-heap-based Any-Gauss-Hit Unit (AGHU) maintains a per-ray hit buffer for efficient volumetric blending of Gaussian primitives. To further optimize traversal performance, GauTracer employs a treelet-based BVH traversal scheme with a far-node pruning scheme, reducing unnecessary node visits beyond the buffer's capacity. Simulation results based on Vulkan-Sim show that GauTracer achieves average performance improvements of 5.8× for 3D Gaussians and 7.3×7.3 \times for 2D Gaussians over the baseline GPU-RTA. It also enhances energy efficiency by 6.6×6.6 \times and 7.5×7.5 \times, respectively.

问问这篇 Paper

问问你的智能体。

Lune 读过与它相关的顶会 Paper,每个回答都会注明依据哪几篇。

可以从这些问题问起

智能体调用

Lunesearch_papers

在 Lune 里问

免费开始,无需绑卡

相关 Paper

黄昏的海面,两侧是细线勾勒的悬崖