Abstract

Rendering large-scale 3D datasets in real-time presents significant challenges due to the widening gap between the ever-growing scale of dataset complexity and the limited memory capacity of modern GPUs. As datasets increasingly exceed available GPU memory, traditional methods encounter severe bottlenecks, such as restricted CPU-GPU bandwidth and inefficient memory usage, which significantly hinder rendering performance. This is especially problematic in interactive applications where high frame rates and low latency are critical. To address these limitations, this dissertation presents UltraRenderer, a novel out-of-core rendering method designed to maximize visual fidelity under strict memory and bandwidth constraints. The core of our approach is a parallel in-place GPU memory management strategy coupled with a coherence-aware level-of-detail selection algorithm. Unlike traditional out-of-place methods, our inplace architecture dynamically reallocates GPU memory slots and consolidates frame-different and reusable data. This strategy conserves nearly half of the GPU memory during runtime, enabling the method to support larger datasets. Furthermore, the method employs an incremental refinement strategy that prioritizes essential data transfer, allowing the rendering quality to converge toward desired fidelity over subsequent frames. Experimental results demonstrate that UltraRenderer significantly outperforms existing state-ofthe- art methods, successfully rendering scenes with billions of triangles and vertices in real-time. By effectively bounding memory usage and optimizing data transfer through frame-to-frame coherence, the proposed method ensures consistent, high-quality rendering performance. The outof- core approach in this method is also demonstrated to support real-time rendering of large-scale 3D Gaussian representations, which use small, colored, semi-transparent ellipsoids (3D Gaussians) to represent a scene rather than polygons. This adaptation validates the broad effectiveness and applicability of our memory management strategies across diverse 3D representations.

Publication Date

3-2026

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Chao Peng

Advisor/Committee Member

Ji Hwan Park

Advisor/Committee Member

Joe Geigel

Campus

RIT – Main Campus

Download

Share

COinS