fVDB v0.5.0 Turns NVIDIA's Library Into Multi-GPU Gaussian Splatting Stack

NVIDIA has shipped fVDB v0.5.0, the release in which its library of GPU takes a significant step forwards. Tagged as the latest release of openvdb/fvdb-core, the drop consolidates a long run of work spanning the v0.3.0 through v0.5.0 changelog into a single coherent pipeline built around one user facing object, fvdb.GaussianSplat3d.

The 0.4 release we covered ahead of GTC is where they added both MCMC densification, and projection via the 3D Gaussian Unscented Transform.

fVDB consolidates into a full stack with gsplat parity and multi-GPU throughout. The MCMC path now has a proper relocation kernel with unit tests and an add-noise kernel, Python bindings and pytests, and a set of exposed knobs that make it usable as a real optimizer. min_opacity surfaced in the relocate function and the k and t parameters exposed in MCMC relocation. Crucially for NVIDIA’s target of enterprise-scale reconstruction, those MCMC kernels gained multi-GPU support through the PrivateUse1 backend, the same partitioning story Sanja Fidler framed as the point of fVDB in our interview with her on reconstructing reality in 3D.

This cycle also added dense rasterization using 3D Gaussians and then an explicit push to improve parity with gsplat for dense rasterization. gsplat is of course the Nerfstudio maintained implementation of gaussian splatting, which has recently seen significant work directly from NVIDIA engineers. Around that core, fVDB now supports backgrounds, masks across all Gaussian render methods, rendering of all contributing Gaussian IDs and weights for downstream analysis, and shared-memory optimizations for the rasterizer.

Rendering was refactored onto a composable camera model for projection and ray generation, with a reworked camera API and world space parity and a camera_fov getter/setter added to fvdb.viz.Scene. Second, the Gaussian splatting autograd and pipeline logic was moved out of C++ and into Python, with an accompanying refactor of the splatting ops into extracted utilities, so the training loop is hackable and composable from Python. Multi-GPU scaling was significantly expanded.

It is available now, on GitHub and PyPI.