Abstract
3D Gaussian Splatting (3DGS) is increasingly popular for 3D reconstructiondue to its superior visual quality and rendering speed. However, 3DGS trainingcurrently occurs on a single GPU, limiting its ability to handlehigh-resolution and large-scale 3D reconstruction tasks due to memoryconstraints. We introduce Grendel, a distributed system designed to partition3DGS parameters and parallelize computation across multiple GPUs. As eachGaussian affects a small, dynamic subset of rendered pixels, Grendel employssparse all-to-all communication to transfer the necessary Gaussians to pixelpartitions and performs dynamic load balancing. Unlike existing 3DGS systemsthat train using one camera view image at a time, Grendel supports batchedtraining with multiple views. We explore various optimization hyperparameterscaling strategies and find that a simple sqrt(batch size) scaling rule ishighly effective. Evaluations using large-scale, high-resolution scenes showthat Grendel enhances rendering quality by scaling up 3DGS parameters acrossmultiple GPUs. On the Rubble dataset, we achieve a test PSNR of 27.28 bydistributing 40.4 million Gaussians across 16 GPUs, compared to a PSNR of 26.28using 11.2 million Gaussians on a single GPU. Grendel is an open-source projectavailable at: https://github.com/nyu-systems/Grendel-GS