Case and point:

.\build\bin\Release\sd.exe -m ..\models\checkpoints\sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 20 --sampling-method euler -H 1024 -W 1024 --seed 42 -p "professional photo of a girl laying down on grass, fine details, 4k resolution" --vae-tiling

I can almost count all the 64 tiles by eye. I don't think that's the expected behavior.

For comparison, here is with normal vae on cpu: