AMD researchers have apparently found a unique way to render the "troublesome" 3D trees, which allows the GPU to significantly improve VRAM utilization.
AMD's "Work Graphs" Technique Reduces GPU VRAM Utilization Massively, Tackling The Troublesome 3D-Tree Generation
With growing graphical computational demands, the industry is in dire need of innovative strategies to improve performance, instead of relying on architectural advancements. With that, AMD researchers have come up with solutions that cater to the highly demanding rendering of 3D tree models. This brings in significant improvements from procedural rendering, to the point that it reduces VRAM consumption from 34.8 GiB to just 51 KiB, which marks a decrease of a whopping 600,000 folds, courtesy of a technique which AMD calls "work graphs", and we'll discuss that next.
AMD researcher's real-time GPU tree generation system uses work graphs (w/ mesh nodes) for procedural tree generation. Without work graphs, the trees in the scene would have required 34.8 GiB of VRAM. With work graphs, only 51 KiB
— Compusemble (@compusemble)
Now, rendering "actual trees" is seen as a benchmark here because it contains complex geometric data, which consumes a significant portion of the VRAM if done through procedural rendering. Now, what AMD researchers have done here is that they have generated trees on the fly, which means that instead of holding all the geometric data, the GPU uses a set of rules to define how a tree is generated.
Then, through "work graphs", individual GPU shaders are assigned iterative processes, creating a graph-like structure of computations. So, in short, the GPU effectively utilizes VRAM through "divide and conquer".
The work graph technique hasn't become mainstream yet, but it seems to deliver promising results. If you are curious about what competitors like NVIDIA to do for such rendering scenarios, well, Team Green employs "mesh shaders" from the Turing architecture, where rendering is divided into a two-stage programmable geometry pipeline. In particular, with AI, NVIDIA has transitioned into "predictive rendering," where low-resolution pixels are upscaled using the power of artificial intelligence. So, it is safe to say that there's a lot going on in the rendering world since we cannot rely solely on computing power.