Except DLSS 5 isn’t just upscaling. It’s replacing the image.

Technically all upscaling replaces the frame with a higher resolution frame.

Even with non-AI upscaling, like linear or bicubic, the original frame isn't copied and then upscaled. The upscaled image is built based on the old image andreplaces the original frame in the frame buffer. DLSS doesn't alter the process, it just uses a neural network instead of a linear/bicubic algorithm.

The new difference with DLSS 5 seems to be that instead of using the frame as the only input it also takes in additional information from earlier in the rendering pipeline (motion vectors) prior to upscaling. This would theoretically create more accurate outputs.

It's kind of like how asking an LLM a question becomes more accurate if you first paste the Wikipedia article which answers your question into the context. Having more information allows for better output quality.

And to achieve what it does, they used one 5090 to render the game normally, and an entire second 5090 just to run DLSS 5.

How is that an improvement in efficiency?

Based on the reporting the use of 2x 5090s in the demo was due to the VRAM requirements of the current iteration, it isn't due to a higher compute requirement. The official DLSS5 release will run on a single card (according to NVIDIA).