A movie-effects tool is heading into chipmaking simulation

The software family that helps artists shape digital clouds is preparing for a different kind of production. OpenVDB grew out of work at DreamWorks Animation and was released as open source in 2012. NVIDIA contributed its GPU-capable relative, NanoVDB, in 2020. Now that family has a proposed job modelling how semiconductor wafers are made.[2]

On 14 October, NVIDIA said Lam Research intends to integrate NanoVDB into SEMulator3D, its semiconductor process modelling platform, with release targeted by the end of 2026. NVIDIA expects up to a 30-fold acceleration for selected wafer-etch modelling. The announced gain is prospective and limited to those modelling tasks; the company has not supplied enough workload and hardware detail to reconstruct it.[1]

The connection is easier to understand by looking past the subject of each simulation. A cloud and a partly etched chip are different objects. Both can create the same computational problem: how to represent complicated three-dimensional information without spending memory and calculation on space that contributes little.

Start with the space you do not need to store

A voxel is the three-dimensional counterpart of a pixel: a small cell that can hold a value, such as a material property. A dense grid reserves a place for every cell. OpenVDB instead uses a hierarchy that stores detailed information where it is needed and represents other regions compactly. The project’s documentation describes both sparse storage and operations on those volumes.[2]

Consider an illustrative grid with 1,000 cells along each of its three axes. It contains one billion cells. At four bytes per scalar value, the values alone require 4 gigabytes in decimal units. If just 1% of the cells need explicitly stored values, that payload would be 40 megabytes. This calculation excludes indexing and other overhead; it illustrates sparsity, not an actual NanoVDB compression result.

The useful insight is that scale depends on what is represented as well as on how fast a processor calculates. Leaving unneeded detail implicit can make a larger problem fit in memory. It can also give an algorithm opportunities to skip regions instead of visiting every cell.

From a process recipe to a virtual wafer

Lam describes SEMulator3D as a physics-driven, voxel-based system. It takes design data and a description of the manufacturing process to build a virtual three-dimensional structure. Engineers can examine how process steps affect the result, investigate defects and explore the range of conditions within which a process works.[3]

That makes geometry part of the engineering question. A manufacturing step can remove material or add it; the shape left by one step becomes the starting point for another. A virtual process flow gives engineers a way to investigate those consequences before committing to a physical wafer experiment. Lam presents this reduction in physical learning cycles as one of the product’s purposes.[3]

NanoVDB changes how the underlying volume can travel through the computer. Ken Museth’s 2021 paper describes a compact, linearised version of the VDB tree held in contiguous memory, without the pointers that would otherwise link separately allocated pieces. That layout helps move the representation between processors and makes it usable on GPUs. The paper demonstrated rendering and simulation applications, including fire and smoke.[4]

There is a specific trade-off. Current NanoVDB documentation says stored grid values can change, while the tree structure locating those values is fixed for a given grid. Changing that structure requires rebuilding it. NanoVDB is therefore a particular data representation with useful constraints; it is not itself a learned AI model.[5]

Where the AI connection becomes real

A separate branch of the family shows why this representation matters to AI. In 2024, NVIDIA researchers described fVDB, which combines NanoVDB with deep-learning operations through a PyTorch extension. Operations such as convolution and attention let neural networks work on sparse three-dimensional data. The authors described applications including reconstructing surfaces from point clouds, generating 3D scenes and adding detail to lower-resolution physics simulations.[6]

Here the shared representation connects two kinds of work: storing and navigating a volume, and learning relationships within it. A neural network still has to be trained for its task. The infrastructure gives it a practical way to handle a complicated spatial domain without first forcing everything into a dense block of equally detailed cells.[6]

Even before that release, the connection is substantial. Film software helped develop an efficient language for three-dimensional space. That language can support a virtual manufacturing process or become the working material of a neural network. The industries share a hard problem that the images on their screens make look deceptively easy.

Report an error or suggest a correction