Mirror-like reflections are not exactly a worst case scenario. Glossy reflections/refractions and GI on the other hand generate more challenging ray distributions.
The impact of the shape of the object is identical with the impact of "roughness", or the impact of more bounces. A skyscraper city of boxes is a lot more performing than a skyscraper city of cylinders or tori. And a city of perfect mirror cylinders can perform identical to a city of blurry glass boxes. Additionally, you'd have to factor in the quality of the BVH, but that is a much worst can of worms really (it's NP-hard).
In general, in raytracing, for performance, what matters is the divergence (or convergence) of directions, because that plays into predictability and locality of memory locations / access. You can practically use the metric of entropy (from compression) to measure ray redundancy, and that redundancy leads to better memory / data access. Try to compress a random data-set, good luck. So, if you pick the perfect rays, you always pay the full trip to memory, because they are truly unique, and perfectly distributed (from an information point of view) thus non-local and unpredictable.
Don't forget that texturing is also at times unpredictable, and an existing bandwidth consumer which can't really be eliminated; or you use cell-shading.
Only if you're in the totally scewed area of very little rays, or highly redundant rays will you find that TFLOPS are useful, because access is not the dominating part.
I guess such demos won't fly in the current_year puritan socio-political climate