Odin: Primitive-Level Synchronization for Distributed Point-Based Neural Rendering

2026-07-22

AI Summary

This work addresses the scalability bottleneck in distributed point-based neural rendering caused by global synchronization barriers, which severely limit training efficiency for large-scale scenes. To overcome this limitation, the authors propose a primitive-level fine-grained synchronization mechanism that replaces conventional global barriers. Their approach employs pre-scheduling to identify low-conflict overlapping computation windows and dynamically releases primitive states through runtime validation, enabling dual-path strategies that prioritize either reconstruction quality or training throughput. The method achieves significant performance gains without compromising visual fidelity: on 8 GPUs, it delivers an average speedup of 1.22× and hides 82% of critical-path waiting time; on 64 GPUs, it outperforms Grendel by up to 1.89× in training throughput.

Abstract

Point-based neural rendering (PBNR) represents 3D scenes as explicit, trainable primitives and underpins high-quality reconstruction and emerging embodied AI and world-model pipelines. Unlike layer-structured neural networks, PBNR has primitive-indexed dependencies: each view reads and updates only a sparse, view-dependent subset of mutable scene state. As large scenes require distributed training and optimized renderers reduce per-view computation, global task- or iteration-level barriers increasingly place synchronization, rather than rendering, on the critical path. We present Odin, a distributed PBNR training system that replaces global barriers with primitive-level synchronization. Its ahead-of-time scheduler uses stable locality and phase order to identify low-conflict overlap windows, while the runtime validates primitive publication before later work observes mutable state. Odin provides a quality-first path that preserves synchronized-training visibility and a throughput-first path that uses overlap and gradient evidence to admit only small, low-impact delayed reads; structural changes and high-impact cases remain synchronized. Across four existing PBNR pipelines and 13 non-city scenes on 8 GPUs, Odin improves throughput by 1.22 times on average and hides 82% of critical-path wait while preserving reconstruction quality. In a MatrixCity mixed-parallel case study scaling to 64 GPUs, Odin improves throughput over Grendel by up to 1.89 times without changing renderer kernels, optimizers, training budgets, or model capacity.

Problem

• distributed rendering
• point-based neural rendering
• synchronization bottleneck
• primitive-level dependencies
• training throughput

Innovation

• primitive-level synchronization
• distributed neural rendering
• point-based neural rendering
• overlap scheduling
• mutable scene state