Modelisation / rendering

Monday, Mar 17 9:00 PM - 9:40 PM CET : An Introduction to OpenUSD [S72488]

Aaron Luk : Director of Product Management, NVIDIA
Slides
Express all data need into USD
UsdMtlX -> Material X for USD
Filling feature gaps (prototyping with custom attribute, and then integrate it into the standard)
Making CAD first class cytizen of USD
Deformable OK in USD
Solidworks does not support USD yet but oops does
Expressing the use of the right solver for soft bodies is a challenge

Tony Kanell : Program Manager, NVIDIA
David Hart : Senior Manager, NVIDIA
Slides
OPTIX 9 from february
A lot of feature will come for Vulkan
Cooperative vectors (shaders have access to the tensor cores now)
Small neural network in shaders (5 to 10 layers, of size 64)
Big enough to do material rendering
Practically only GEMM
Stack Vector matrix multiply -> matrix matrix multiply if you have many vectors
Remap vectors to matrices efficently
Enable partial warps
Provide a more standardized interface than tensor cores
Textures 20x smaller than compress DDS
RTX mega geometry ->bottleneck is BVH (only 5 million triangles per frame)
Clusters take advantage of meshes repetitions or rendundancies
BVH build is similar to ease a binary search
How can we reuse the topology in the BVH
faster to refit than to rebuild
Predefined grid template that are used by everyone
Similar to nanite meshes
The time to build the BVH is 100x slower than the rendering time -> need for optimisation
Subdivition surfaces tessellation library using Clusters (OSD lite base of open sub div)
100 millions triangles at 60fps
DLSS on top of that
up to 50x speed up
Rocaps : swept sphere (fine tuning is available in optix kernel). Only for OptiX now. Not Vulkan or directX
LSS : Linear Swept Sphere => now Hardware accelerated
RTXCR : character rendering, based on Vulkan
Sphere are also Hardware accelerated
OptiX headers released on Github
Clusters and Cooperative vectors are available on Turing (T4, etc)