VkSplat: High-Performance 3DGS Training in Vulkan Compute

TL;DR

VkSplat introduces a Vulkan-based 3D Gaussian Splatting training pipeline that significantly improves speed and VRAM efficiency while ensuring cross-vendor compatibility.

Contribution

It is the first fully Vulkan-based 3DGS training pipeline achieving state-of-the-art performance and broad GPU compatibility.

Findings

Achieves 3.3x speedup over CUDA+PyTorch baseline.

Reduces VRAM usage by 33% compared to existing methods.

Demonstrates compatibility across multiple GPU vendors.

Abstract

We present VkSplat, a high-performance, cross-vendor 3D Gaussian Splatting (3DGS) training pipeline implemented fully in Vulkan compute, addressing performance and compatibility limitation of existing training pipelines. With various optimizations, we achieve $3.3\times$ speed and $33\%$ VRAM reduction over CUDA+PyTorch baseline, maintaining quality, and demonstrating compatibility across GPU vendors. To the best of our knowledge, this is the first fully-Vulkan-based 3DGS training pipeline that achieves state-of-the-art performance. Code: \href{https://github.com/harry7557558/vksplat}{https://github.com/harry7557558/vksplat}