The application of machine learning techniques to 3D Gaussian Splatting (3DGS) has enabled high-fidelity interactive rendering of complex 3D scenes on modest compute resources, offering significant potential for scientific visualization. However, the memory demands of reconstructing large-scale scientific volumes at high fidelity exceed the capacity of individual GPUs. While distributed 3DGS frameworks exist for urban scenes, they rely on global scene analysis or frequent all-to-all communication, inherently limiting their scalability on High-Performance Computing (HPC) clusters. In this paper, we present a scalable distributed 3DGS training and rendering framework designed for large-scale scientific volumetric data. Leveraging the native domain decomposition of HPC simulations, our approach treats each spatial partition as an independent optimization task that avoids device-to-device communication. We further exploit temporal coherence in simulation data to accelerate training on subsequent timesteps. We perform experimental studies demonstrating that our approach achieves near-ideal weak scaling by eliminating synchronization overhead. Additionally, we evaluate the trade-offs of the standard 3DGS optimization strategy against a Markov Chain Monte Carlo (MCMC) approach. We find that while standard 3DGS produces higher quality reconstructions with larger Gaussian footprints, MCMC effectively bounds the primitive count at scale in exchange for a minor reduction in fidelity. Furthermore, we demonstrate that our temporal fine-tuning strategy for time-series data remains robust at scale, establishing a reliable path towards in situ applications of 3DGS for massive scientific simulations.
Sewell et al. (Mon,) studied this question.