The bottleneck for "ray traversal" for regular SIMD GPUs is the intersection testing, ie, ray-box or ray-triangle. There is a repeated misunderstanding from some in the tech press that RDNA2 lacks ray traversal acceleration. It stems from the marketing points from NVIDIA. https://twitter.com/Dachsjaeger/status/1390951024286916609

Any acceleration of ray-box or ray-triangle is already meeting the definition of accelerating ray tracing or traversal into the scene, as compared to GPUs which lack this intersection acceleration, ie, prior generations.

It may well be that Ampere is much faster than RDNA2 when it comes to ray-tri testing, and this is why HairWorks with its numerous tiny triangles will negatively impact RDNA2 RT more. But it is wrong to say RDNA2 lacks ray traversal acceleration.

However, dev optimization should avoid situations like that, which they already do, by generating a simpler geometric acceleration structure of the scene. It seems HairWorks somehow bypassed this step, with excessive small triangles choking ray-tri testing.