A workload made of fine structures
Hair is not simply a dense triangle mesh. Large numbers of fine strands create difficult visibility and execution patterns. Grouping those strands and choosing how the GPU processes them can substantially change the shape of the workload.
Clustering and execution
That work included capsule clustering and comparisons of global-pixel and groupshared execution paths. We measured how the organisation and size of the workload affected GPU execution time.
This is implementation-level research: a useful answer needs the data representation, GPU path and measurement conditions together. It also needs to preserve the visual behaviour that the application requires.