UE5 Nanite in WebGPU
An experimental WebGPU project recreating Unreal Engine 5’s Nanite-style “virtualized geometry” sparks both excitement over its ability to render massive triangle counts in the browser and criticism of its technical compromises and naming. Commenters dig into how Nanite-like software rasterization on the GPU can outperform traditional hardware for tiny triangles, why 64‑bit atomics and robust feature detection are crucial yet poorly supported on many platforms, and how this approach compares to more conventional LOD systems in engines like Godot, Unity, and Bevy. Broader points focus on the maturity of WebGPU across browsers, the gap between research prototypes and production readiness, and the long-term prospect of dedicated hardware support for virtual geometry.
WebGPU support and demo reliability
- Many users report the demos failing despite enabling WebGPU (especially on Safari and Linux/Chrome, some Android and Windows setups).
- Errors include missing functions in shaders, misaligned buffer sizes, and device loss on Windows.
- Some see a generic “No WebGPU available. Please use Chrome” message even when WebGPU is working elsewhere.
- A subset of users report good performance (e.g., 60–120 FPS on M-series Macs), indicating environment- and driver-specific issues.
64‑bit atomics and visibility buffer design
- Nanite-style software rasterization relies on 64‑bit atomics to pack depth and other per-pixel data into a single 64‑bit value and do atomic min/max.
- WebGPU’s lack of 64‑bit atomics forces compromises: reduced depth precision, minimal material data, and visible artifacts.
- Hardware rasterizers have fixed-function logic doing this depth/visibility work; Nanite moves it into compute.
Why software rasterization can beat hardware for Nanite
- GPUs shade in 2×2 pixel quads; very small triangles waste work because many shaded pixels lie outside the triangle.
- Nanite-like approaches use compute shaders (“software rasterization on the GPU”) for tiny triangles and fall back to hardware rasterization for larger ones.
- In scenes dominated by near-pixel-sized triangles, compute rasterization can be faster; for larger triangles, it’s slower, so a hybrid heuristic is used.
- There is disagreement on overall cost: some see Nanite as significantly slower than good traditional LOD for typical triangles; others emphasize its gains in small-triangle regimes.
Nanite / virtual geometry concepts and ecosystem
- Nanite is described as a bundle of techniques: meshlet-based LOD hierarchies, aggressive compression and quantization, per-cluster dynamic LOD, software rasterization, and streaming.
- Core idea: maintain roughly pixel-level geometric precision and only process the detail actually visible on screen.
- Similar “virtual geometry” systems exist in other engines (e.g., Bevy); research roots date back at least to a 2009 dynamic mesh simplification dissertation.
- Some argue the GitHub project name is confusing/trademark-adjacent, since it’s an independent reimplementation rather than UE5’s Nanite itself.
Usability and controls
- Users praise the technical achievement and visual density but note awkward camera controls on mobile and touchpads.
- Suggestions include using device orientation APIs for more intuitive mobile interaction.