Didn't see a thread on this but apparently there have been some misconceptions regarding the PS5 Lumen in the Land of Nanite demo, suggesting it can only run on PS5. Brian Karis decided to put those rumors to bed:
Full livestream here:
Some notes from the Land of Nanite portion of the livestream:
And finally, a note from Karis:
Full livestream here:
Some notes from the Land of Nanite portion of the livestream:
- The goal (dream) with Nanite was to remove the notion of technical budgets and constraints for 3D Artists so that they can just focus the art itself, regardless of quality
- The virtualization of geometry is a lot harder to implement in practice than the virtualization of textures due to factors like rendering cost
- GPU driven pipeline for Nanite, which allows optimizations like reduced draw call requirements per scene
- Draw calls are per material, not per object. All opaque geometry can now be rendered with just a single draw call
- In this demo, pixel sized triangles used software rasterizer, as it's up to 3x faster than hardware rasterizer. Big triangles use hw rasterizer (rasterizer chosen on per cluster basis)
- Average frame for demo was ~1400p upscaled to 4k via temporal upsampling
- Nearly zero CPU time and very little CPU cost. Most of the work handled by the GPU
- Proprietary compression tech to keep file sizes in check (the immense amount of geometry would take up too much space otherwise)
- All Nanite data on disk (compressed) was 6.14 GB in total
- Texture data far larger than geometry data
- 16k x 16k virtualized shadow maps work in a similar way to Nanite by only rendering shadow detail that's perceptible (shadow pixels)
- The scenes shown are demonstrated using a free roam camera to show that scripting is not necessary and that there isn't any masked loading
- Entire level was loaded into memory (mainly for the purposes of showing that the infamous crack was not used to hide streaming)
- Self-shadowing makes it easier to see the increased fidelity of actual geometry compared to normal maps
- Demo shown is not using world partitioning (the tech hadn't fully developed by the time the demo was made)
- The closer the camera is to a given mesh, the more faithful the geometry will be relative to the source model (the delta in fidelity is practically imperceptible without visualization data)
- 128 triangles per cluster. Cluster sizes on screen are roughly the same, regardless of viewing distance, which allows the triangles to have similarly stable counts on screen
- On average, there weren't much more than 20M tris rendered on screen, even with billions of triangles from source geometry. The tri count will scale with resolution
- Nanite supports hard surface model meshes as well, not just organic assets from megascans/photogrammetry
- The ring on the portal at the end of the demo is not actually a unique mesh, but made from instanced geometry like the statues shown earlier in the demo
And finally, a note from Karis:
