Have written elsewhere in this forum that Nvidia would have to cut RT & Tensor cores, integrating their functionality in CUDA cores, following the RDNA2 design (used in consoles also).
The Turing dies are too big to to shrink at 7nm EUV with RT & Tensor cores (let alone 5nm), nor are compatible with the Nvidia announcements about MCM gpus in the future. Which means any dedicated RT cores usage is going to be short lived like the dedicated PGUs were with Physx.
Personally I wouldn't be surprised if Ampere is still at 12nm and not 7nm EUV as is rumoured, if Nvidia maintains the RT & Tensor cores until moved to unified CUDA with MCM design at 5nm.