Odin speeds up distributed neural rendering by 1.89x on 64 GPUs

Point-based neural rendering (PBNR) is central to high-quality 3D reconstruction and emerging embodied AI pipelines, but training large scenes across multiple GPUs suffers from synchronization overhead. Global barriers force all GPUs to wait for the slowest worker, even though each view only touches a sparse subset of scene primitives. Odin, a new system from researchers at Tsinghua University and other institutions, tackles this by replacing those barriers with primitive-level synchronization designed specifically for PBNR's access patterns.

The key innovation is a two-pronged approach: an ahead-of-time scheduler that identifies low-conflict overlap windows using stable spatial locality and training phase order, and a runtime that validates publication of mutated primitives before other GPUs read them. Odin offers two execution modes — a quality-first path that preserves synchronized-training visibility, and a throughput-first path that allows small, low-impact delayed reads while keeping structural changes synchronized. Tested on four existing PBNR pipelines with 13 non-city scenes on 8 GPUs, Odin improves throughput 1.22x on average and hides 82% of critical-path wait time. In a MatrixCity mixed-parallel case study scaling to 64 GPUs, Odin outperforms Grendel by up to 1.89x without any changes to renderer kernels, optimizers, training budgets, or model capacity.