Path Tracing of Massive Scenes in Blender on Multi-GPU Systems
, VSB-Technical University of Ostrava. IT4Innovations National Supercomputing Centre
Learn how to minimize the performance loss of path tracing when working with distributed scene data through a methodology built on memory management. Besides conserving high performance, it allows rendering of massive scenes that are larger than the memory capacity of a single GPU in the multi-GPU system. It takes full advantage of the NVLink interconnect that enables up to 16 GPUs to share their memory. The main concept is that those parts of the scene data that have the highest amount of memory accesses are duplicated on all GPUs to reduce negative impacts on performance. Proof of concept will be shown using Blender Cycles path tracer. We show rendering of several large scenes, such as the Moana Island scene with a size of 169 GB.