I am assuming this is still the case (that they are not doing too much work in the gaming cards ) but i was talking about this particular case where it is known the tensor cores were not used. It was still good enough performance. And yet some people are convinced it can't be done without the tensor cores, even if most of them have used DLSS done without tensor cores already and they were impressed at that time.Quite a bit of stuff in the DLSS and RT pipelines don't use Tensor cores (even though originally claimed was the case) even though in a best case scenario they can provide a significant performance boost over shaders due to the lack of flexibility and having to hand tune to avoid penalities i.e. Tensor processing is often last in the chain and can't be run simultaneous with other work.
When and where you can use Tensor cores though the speed up can be significant.
At lower resolutions you could probably get away with using Shaders but towards 4+K resolutions you would have to rely on Tensor cores to get decent performance gains.
Off the top of my head IIRC doing it on the shaders on Ampere takes about 4ms at 1080p compared to ~1ms on the Tensor cores at 1440p (output resolution) - I don't know the numbers for 1080p on Tensor but I'd assume sub 1ms.