This article is available in: Français (French)
At SIGGRAPH 2022, NVIDIA unveiled NeuralVDB, a new technology aimed at “bringing AI and GPU optimization to OpenVDB”.
As a reminder, OpenVDB is an open-source C++ library widely used in the VFX and animation industry for simulating and rendering sparse volumetric data. OpenVDB provides both a hierarchical data structure and a suite of tools in order to store and manipulate this data. To sum it up: OpenVDB is a key component of the way we simulate and render water, fire, smoke.
NeuralVDB is introduced by NVIDIA as an improvement over OpenVDB, with impressive features. According to NVIDIA, “NeuralVDB delivers up to 100x reduction in memory footprint”. This dramatic improvement could be a game changer, since complex volumetric datasets require a huge amount of memory. NeuralVDB could allow users to work on the same projects on lower end workstations.
Furthermore, NeuralVDB also includes the GPU acceleration of NanoVDB, which was unveiled last year by NVIDIA. According to the company, NanoVDB “opens the door to real-time simulation and rendering”.
NeuralVDB is therefore both faster and lighter thant OpenVDB. It also allows the weights of a frame to be used for the following one, which accelerates training by up to 200%. Furthermore, by using the network results from the previous frame, explains NVIDIA, NeuralVDB “enables users to achieve temporal coherency, or smooth encoding”. One of the consequences is that users won’t have to rely as much on post-processes such as motion blur to get the result they want.
NVIDIA announced that NeuralVDB was “built to be compatible with preexisting VDB pipelines”: adoption should therefore be quite seamless. NeuralVDB will “soon” be available in beta: in the meantime, a form on the official NeuralVDB webpage will allow you to register and get notified of upcoming updates.