cs.DCEmpirical
Odin: Primitive-Level Synchronization for Distributed Point-Based Neural Rendering
Jul 22, 2026
The paper presents Odin, a distributed training system for point-based neural rendering (PBNR) that replaces global barriers with primitive-level synchronization, improving throughput and hiding critical-path wait.
The paper introduces Odin, a distributed training system for PBNR that uses primitive-level synchronization to improve throughput and hide critical-path wait.
Keywords
Before reading this…
Applications
To understand this paper, make sure you know these concepts first:
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.
SiFAR: Synchronization-Free All-Reduce for Low-Latency LLM Inference
This paper proposes Synchronization-Free All-Reduce (SiFAR) to reduce All-Reduce…
RL Post-Training Builds Compositional Reasoning Strategies
This paper investigates whether RL enhances existing skills or creates new strat…
DRIFT: Direct Reduced Fourier Transforms for Distributed Spectral Neural Operators
This paper introduces the Distributed Truncated Spectral Transform (DTST) for Fo…
Online Neural Space Time Memory for Dynamic Novel View Synthesis
The paper proposes a method for online novel view synthesis from multi-view stre…
GrainGS: Gradient-Decoupled Gaussian Splatting for Efficient Dynamic Novel View Synthesis
The paper presents GrainGS, a dynamic Gaussian framework for scene reconstructio…
Joint Synchronization and Sensing in Networked ISAC via Structured Canonical Polyadic Decomposition
This paper proposes a structured canonical polyadic decomposition (SCPD) algorit…
Toward Semantic Communication for Real-time Mobile 3D Reconstruction
This paper proposes a SemCom framework for real-time mobile 3D reconstruction, w…
Rendering 3D Gaussians on a Graph Processor
First implementation of a 3D Gaussian renderer on an Intelligence Processing Uni…