- The paper demonstrates that leveraging Vulkan compute significantly reduces VRAM usage by 60% while maintaining robust 3DGS reconstruction quality.
- The methodology employs advanced buffer pre-allocation and both adaptive and fixed-size MCMC densification to optimize training across NVIDIA and AMD GPUs.
- Empirical results reveal competitive PSNR and SSIM metrics alongside clear vendor-specific performance gaps, guiding future cross-platform neural rendering research.
Overview
VkSplat introduces a Vulkan-based compute pipeline for efficient training of 3D Gaussian Splatting (3DGS) models. The approach emphasizes high-performance execution on both NVIDIA and AMD consumer GPUs, contrasting with traditional CUDA/Torch implementations such as GSplat. The study assesses VkSplat across seven scenes from the Mip-NeRF 360 dataset, employing rigorous benchmarking with strict reproducibility standards.
Methodological Innovations
VkSplat leverages Vulkan's explicit memory and compute control, facilitating fine-grained optimization unattainable with existing PyTorch/CUDA abstractions. The pipeline pre-allocates buffers where feasible, minimizes VRAM spikes via controlled buffer resizing, and exploits advanced parallel primitives for sorting and rasterization. Two densification methods are evaluated: (1) default adaptive densification tailored to the content, and (2) fixed-size MCMC-based densification with up to one million Gaussians. The VRAM accounting methodology distinguishes between total buffer allocations (for fair academic comparison) and peak usage (for practical OOM avoidance), addressing inconsistencies in PyTorch reporting versus actual hardware measurements.
Empirical Results
Quality Metrics
VkSplat achieves robust reconstruction quality: average PSNR of 29.2 and SSIM of 0.88, with LPIPS Alex averaging 0.125 under default densification. MCMC densification provides similar perceptual quality, albeit with marginal increases in LPIPS values and minor PSNR/SSIM fluctuations. Both densification schemes demonstrate strong fidelity across diverse indoor/outdoor scenes.
Resource Utilization
A key empirical claim is that VkSplat reduces VRAM consumption by approximately 60% compared to GSplat on identical hardware, supported by direct nvidia-smi polling. GSplat's reported VRAM systematically underestimates actual usage, whereas VkSplat's numbers are closer to real-world consumption, with only an 18% underreport, versus GSplat's 38%. This demonstrates the practical advantage of VkSplat for memory-constrained environments and scalable multi-GPU setups.
Timing Analysis
VkSplat delivers competitive training times on NVIDIA RTX 3090 under both densification protocols. Pipeline stage breakdown reveals that Vulkan's compute benefits are most pronounced in rasterization and optimization steps. On AMD Radeon RX 7800 XT, VkSplat maintains high computational throughput, but is bottlenecked by PCIe-based image transfers, which are nearly 30× slower than on NVIDIA hardware. The paper postulates that asynchronous data transfer could mitigate this, proposing future optimization avenues.
Numerical Highlights
- Peak VRAM for VkSplat (default densification) averages around 3.06 GiB versus GSplat's >8.68 GiB reported (actual 11.98 GiB), highlighting a substantial memory footprint reduction.
- MCMC densification achieves peak VRAM near 0.92 GiB, supporting ultra-efficient 3DGS fits even at one million Gaussians.
- AMD/NVIDIA timing breakdowns reveal clear pipeline-stage disparities, directly informing cross-vendor optimization strategies.
Contradictory and Bold Claims
VkSplat asserts that VRAM reductions are not artifacts of unfair accounting, substantiated by side-by-side nvidia-smi polling and software-level VRAM tracking. The empirical evidence indicates that Vulkan compute yields genuine, measurable gains in resource efficiency and practical usability, especially when compared to PyTorch/CUDA-based implementations.
Additionally, VkSplat demonstrates that fixed-size densification (MCMC) achieves comparable or better peak VRAM behavior thanks to buffer pre-allocation, diverging from existing GSplat practices where buffer spikes frequently trigger OOM failures.
Implications and Future Directions
VkSplat's results suggest that Vulkan compute is a viable and scalable alternative for 3DGS training pipelines, especially in heterogeneous, consumer-grade GPU environments. The explicit memory control and vendor-agnostic compute primitives enable broader accessibility for researchers and practitioners. These findings imply that future AI frameworks may increasingly leverage Vulkan for high-throughput, cross-platform training and inference, especially as CUDA dominance wanes and AMD hardware adoption grows.
Practical implications include the potential for multi-scene, large-scale 3DGS optimization on devices with limited VRAM, and seamless extension to real-time applications such as SLAM, AR/VR, and photogrammetry. The theoretical ramifications center on the decoupling of high-performance 3DGS from proprietary CUDA/Torch ecosystems, advocating for open, cross-vendor standards in vision-centric neural rendering.
Anticipated future developments encompass further elimination of PCIe transfer bottlenecks via DMA/asynchronous techniques, more granular Vulkan-specific rasterization optimizations, and integration of Vulkan compute with AI-centric platforms such as ONNX Runtime or VulkanML.
Conclusion
VkSplat establishes a clear benchmark for high-performance, low-memory 3DGS training via Vulkan compute, substantiating its empirical claims against prevailing CUDA/Torch alternatives. The pipeline's efficient resource utilization and extensible design herald a shift toward vendor-agnostic, scalable neural rendering frameworks. The work's results invite further exploration of Vulkan's role in AI research infrastructure, both for immediate practical deployment and as a foundation for future cross-platform, high-performance training paradigms.
