New architectures impact on optical flow SDK's performance/capabilities

Optical flow SDK was introduced with Turing architecture; I understood that it relies on specific hardware capabilities.

However, the subsequent architecture white papers did not develop the impact on the optical flow capabilites.

What are the impacts (qualitative and quantitative) of Ampere and Hopper/Lovelace on the optical flow through the SDK ? Is there any documentation about it somewhere ? If not, how can I estimate the compute power evolution from Turing to the following architectures and chips (eg., based on FP32 cuda cores count or something) ?