NeRArch-Sim: A Unified Simulator for Benchmarking and DSE of Neural Rendering Accelerators

Recent breakthroughs in neural rendering promise numerous real-time 3D intelligence applications, e.g., AR/VR, robotics, and digital twins. To satisfy real-time requirements, various neural rendering accelerators have been developed, with each employing a different rendering algorithm pipeline and designed for a specific hardware target. Such one-off specialization, representing only a single design point, creates two gaps: (i) fair, extensible cross-accelerator benchmarking, and (ii) design-space exploration (DSE) that retargets designs across pipelines or hardware budgets. We present NeRArch-Sim, an open-source simulator purpose-built for neural rendering accelerators with module-accurate models and validated end-to-end schedules (error $\leq 9.4 \%)$. NeRArch-Sim follows two principles: (1) modular abstractions for software workflows and hardware, enabling extensible benchmarking across diverse algorithmic pipelines and accelerator designs; and (2) a dataflow-aware two-level scheduler for rapid, effective DSE. Across eleven accelerators and two datasets, NeRArch-Sim reproduces prior designs with minimal changes (modeling error $\leq \mathbf{9. 4 \%})$ and guides new accelerators that achieve up to $1.3 \times$ efficiency gains. To our knowledge, NeRArch-Sim is the first open-source simulator that models a wide range of neural rendering accelerators, providing timely infrastructure for this emerging domain. The simulator is publicly available at https://github.com/GATECH-EIC/NeRArch-Sim.