I've been sceptical of the rumours up until now but 450W for the 4090 is much more believable, that combined with 15% of the cuda cores being left on the table so that there is room for both a 3090TI and a Titan is very NVIDA.
One thing that not many people have mentioned is that while the rumoured FP32 performance is going to more or less double, the ram bandwith is only increasing by about 10%. On the surface you might think that could limit performance significantly but i have a feeling that it wont be all that bad. Looking at traditional Raster based games my guess is that it probably wont throttle performance at all, in God of War for example the memory bus utilization on a 3090 is only around 35%, so that leaves plenty of scope for more performance. On the other hand in ray traced games regularly see well above 50% bus utilisation, so i wouldnt be completely surprised if we see some strange scaling vs Ampere in RT games. That being said NVIDIA have significantly increased the on die cache for Lovelace similar to infinity cache for AMD on the 6900 series which make do just fine with a 256bit bus and 18GB/s memory. It's not unreasonable to think similar is possible for NVIDIA.