NVIDIA describes fVDB as a deep learning platform for "sparse, large-scale, and high-performance spatial analysis," enabling software developers to create "AI architectures and algorithms scalable to real-world sizes." It utilizes OpenVDB, an open standard for volumetric data, and NanoVDB, a simplified representation of this data on NVIDIA GPUs, as an efficient way to handle such large-scale datasets. fVDB then builds AI operators on top of NanoVDB, allowing for common tasks such as convolution, pooling, attention, and mesh generation.

The primary practical use case for fVDB appears to be the creation of large-scale digital twins of the real world for applications such as urban planning and industrial modeling. However, there are also clear potential applications in the entertainment sector, including creating computer-generated cities for visual effects or game development.
The video at the beginning of the news showcases some intriguing examples created using fVDB, including a generative AI model for producing voxel representations of entire city blocks at low resolution, complete with buildings, roads, and trees. Another demonstration shows fVDB being used to update an existing facial model. NVIDIA told CG Channel that fVDB is a "very general technology," and they anticipate it will be used extensively.

NVIDIA positions fVDB as a more powerful solution compared to existing similar technologies, capable of handling data at "4 times the spatial scale and 3.5 times faster than previous platforms." It is also designed to simplify implementation, offering "user-friendly APIs so you don't have to merge different libraries."
fVDB is available as a PyTorch extension, can read and write existing VDB datasets "out of the box," and integrates with existing NVIDIA technologies such as Warp and Kaolin.

The source code for fVDB will be available as part of the GitHub repository for OpenVDB, supported by the Academy Software Foundation. The functionality of fVDB will also be accessible through NIM, NVIDIA's container system for deploying GPU-accelerated AI microservices. NVIDIA plans to launch three fVDB microservices for mesh creation, large-scale NeRF generation in USD, and enhanced physical simulation.

Any applications built on fVDB will need to run on NVIDIA hardware. The framework is "built from the ground up on core NVIDIA technologies," including CUDA, NVIDIA's proprietary GPU computing API, and Tensor Cores in modern NVIDIA GPUs. The fVDB database is expected to be merged with the OpenVDB GitHub repository "in the near future." The source code in the repository will be available under the open-source MPL 2.0 license. Software developers can also apply for early access to the fVDB PyTorch extension. NVIDIA has not yet announced a release date for the fVDB NIM microservices.