VkSplat: High-Performance 3DGS Training in Vulkan Compute

VkSplat, introduced by researchers from (presumably academic institutions), is a high-performance training pipeline for 3D Gaussian Splatting (3DGS) implemented entirely in Vulkan compute shaders. Previous 3DGS implementations relied heavily on CUDA and PyTorch, limiting training to NVIDIA GPUs and often suffering from vendor-specific bottlenecks. VkSplat leverages Vulkan, a cross-platform, cross-vendor GPU compute API, to achieve state-of-the-art performance without proprietary lock-in. The pipeline incorporates multiple optimizations, including fused kernel designs, memory pooling, and efficient sorting, resulting in a 3.3x training speedup and 33% reduction in VRAM usage compared to the standard CUDA+PyTorch baseline. The authors report no degradation in rendering quality, matching or exceeding baseline PSNR scores on standard benchmarks.

The significance of VkSplat extends beyond raw performance. By targeting Vulkan, it works on GPUs from NVIDIA, AMD, and Intel, democratizing 3DGS training for developers and researchers without access to high-end NVIDIA hardware. It also opens the door to embedded and mobile deployments where CUDA is unavailable. The project is open-source and submitted to Eurographics 2026. For professionals in computer graphics, AR/VR, and autonomous systems, VkSplat offers a faster, memory-efficient path to real-time 3D reconstruction, while reducing dependence on a single GPU vendor. This could accelerate adoption of 3DGS in applications like spatial computing, digital twins, and photorealistic rendering.