GTC 2020: Next-Gen Rendering Technology at Pixar

GTC 2020 S22266
Presenters: Max Liani,Pixar
Abstract
We’re continuing on the journey to develop RenderMan XPU, our next-generation photorealistic production rendering technology to deliver the animation and film visual effects of tomorrow. We’ll update you on our techniques leveraging heterogeneous compute across CPUs and GPUs, how we adopted the unique features offered by NVIDIA RTX hardware, and how we combine that with our own solutions to make RenderMan the path tracer of choice for so many productions.

Watch this session
Join in the conversation below.

The term “Sort” seems to be the wrong term to use here, leading to confusion:
The open problem discussed here is about bunching “like-things” together in time and space (sequential in memory to improve coherency, and executed together to reduce context-switching latency), countering the random nature of path tracing (random in terms of types of rays, geometries materials, etc.).
Sorting is about “order” and dependencies, not about “categories”.
In virtually all the cases discussed here, the order within each category within a pipeline-step, is irrelevant - there are no dependencies - each ray/computation is independent of the others within each category, and even between categories in a pipeline step.
Literally any instance of the term "Sorty by " in the slides would be better substituted by the term "Group by ".
So it’s not a “Sorting” problem of the kind that transparency is in rasterisation, it’s a “Grouping” problem, so the comparison furthers the confusion. The ordering in rasterisation is a bout “correctness”, whereas the grouping in path-tracing is about performance. The problems are not related.