Traditionally, NVIDIA’s graphics architectures have been superior to rival AMD parts at rendering complex geometry including tessellation and meshes with extra high polygon counts. The reasons behind it are multiple, but the crux of it is that GeForce GPUs have dedicated tessellation units and have better driver-level optimizations for the same. Although AMD has made significant progress on both fronts, there’s still a fairly wide gap between the two competitors:

The 3DMark mesh shader benchmark has just been released, allowing us to test the performance of different graphics cards under heavy geometric workloads. As you can see in the above benchmark, the Radeon RX 6900 XT is slower than both the NVIDIA RTTX 3070 as well as the 3080 when running the mesh shader benchmark without the optimization disable. While the Ampere cards average in the late 60s, the RX 6900 XT scores just over 30 FPS.

We did a bit of checking and it turns out that the reason the lower score isn’t limited to the lack of fixed-function hardware units but is also due to core clock throttling.

In the above graph, you can see that the 6900 XT drops down to just over 1GHz for the majority of the first part of the benchmark. The clocks are much more consistent with mesh shading enabled, meaning that the throttling is by design and not a bug.
It would seem that AMD cards throttle when the geometry load on the vector units crosses a certain limit (to avoid slowing down other concurrent workloads?). The same behavior isn’t observed on rival NVIDIA hardware and would explain why the Navi GPUs are still a fair bit slower than competing RTX graphics cards in geometry heavy workloads.
Even with mesh shading enabled and the clock speeds running at spec, the RX 6900 XT is a bit slower than the lower-priced RTX 3080. Granted, the delta isn’t large, but considering that the latter is priced a fair lower than the former, it shows just how some games run much better on a particular hardware architecture compared to the rest.