NVIDIA

NVIDIA has built accelerated computing hardware since 1993 and now sits at both ends of the radiance field pipeline, supplying the GPUs that nearly all training runs on and publishing the reconstruction code that runs on them. Instant NGP established that position in 2022 by collapsing NeRF training from hours to seconds with a multiresolution hash encoding, and Instant NGP 2.0 pushed rendering into real time by fusing the pipeline with just in time compilation through tiny-cuda-nn 2.0. Instant NGP ships under the NVIDIA Source Code License, which requires a separate agreement for commercial use.

3DGRUT is the current research vehicle. NVIDIA committed to open sourcing 3DGRT and 3DGUT in January 2025 and shipped them together under Apache 2.0, pairing an OptiX ray tracer that handles secondary rays, reflections and refraction with a CUDA rasterizer whose unscented transform renders distorted and rolling shutter cameras a standard projection cannot. Version 1.1.0 added a NuRec export pipeline, PPISP and training upgrades, and the repository reads NCore v4 datasets. An in-tree 2.0.0 landed on the main branch in July 2026, generalising both renderers from fixed three channel spherical harmonics to arbitrary dimension per particle features decoded by a small MLP, and adding FP16 storage options and tiny-cuda-nn bindings. NVIDIA has published no v2.0.0 tag and no GitHub Release, so the newest tagged build is still v1.1.0.

fVDB is the sparse spatial data library underneath much of that work. The 0.4 release landed ahead of GTC 2026 and 0.5.0 turned it into a multi-GPU Gaussian splatting stack that distributes a single reconstruction across several devices. On the viewing side NVIDIA maintains vkSplatting, an Apache 2.0 Vulkan viewer with rasterization, ray tracing and hybrid pipelines that shipped as Vulkan Gaussian Splatting 2026.1, moved to 2026.2, which replaced Phong shading with GGX physically based materials and added path traced lighting, and reached 2026.2.9 on September 3 2026 with a smaller ray payload, a fix for splat materials being read from the wrong splat set, portable project files at version 8 and a build that runs on AMD hardware under the RADV driver. It is a viewer only and does not train.

NuRec is the productised path. It is NVIDIA's neural reconstruction library for simulation, it reached general availability in Omniverse and ships inside Isaac Sim 6.0, where a reconstructed real environment becomes somewhere a robot policy can be trained. NVIDIA also formally released NCore, and Alpamayo 1.5 carries the same reconstruction work into autonomous driving. 3DGRUT exports to NuRec directly, which is what connects the research repository to the simulation products.

The practical constraint across the training and simulation stack is that it is NVIDIA only, since 3DGRUT requires CUDA and OptiX, fVDB requires CUDA, and NuRec runs on NVIDIA hardware by design. vkSplatting is the one component that travels, with NVIDIA listing any Vulkan 1.4 capable GPU as a requirement and the 2026.2.9 notes stating that the default build runs on AMD under RADV with DLSS disabled at runtime. The research does travel outward. Nerfstudio integrated 3DGUT as a rasterizer option, Spatial Fields carries 3DGUT scenes into Vision Pro, and third party projects reuse the Apache 2.0 3DGRUT kernels directly, including a Blender fork that ray traces splats in Cycles.