Currently Online 
The only "evidence" shared that DLSS burdens the CPU is what looks to be a quote that describes how the CPU can bottleneck DLSS (especially frame generation) workloads.
A CPU working on a separate part of the pipeline acting as a bottleneck on another part of the pipeline is not the same thing as that part of the pipeline using the CPU for compute.
DLSS upscalers are relatively small CNN or transformer models. The only part the CPU takes is loading the model from the driver located in system storage into the GPU's VRAM.
That likely happens in milli-seconds after you change your game settings. And it only happens as often as you change the settings. All actual inference steps are done on the GPU.
What the quote shared is saying is that like with any other setting that increases framerate (like say if you decrease the resolution to 720p without even using upscaling) there is a higher burden on the CPU from rendering more frames and the CPU can be the bottleneck, limiting GPU utilization and affecting the overall ability to attain higher framerates.
Additionally, there is the issue that DLSS FG works best when input framerates are higher, so there is an incentive to have your input framerate be relatively high, which also burdens the CPU more than otherwise (like if you locked to 30fps and didn't use FG.)
Now it is possible to have a CNN/transformer model utilize both the GPU and CPU, but the only time you would want to do that is when memory is split between the CPU and GPU, the model is too big to load on the GPU and you don't care about the order of magnitude increase in latency. None of this applies to DLSS or the Switch 2.
DLSS models (especially CNN) are relatively small, the Switch 2 has unified memory.
