Excessive VRAM Usage with the XL Model

(ps; The VRAM usage returns to normal when switching to the SD 1.5 model.)
The graphics card is a P40, and the operation doesn't involve using any LORA or high-definition enlargement nodes. It only performs the most basic base+refiner operations at a resolution of 1024. The VRAM usage reaches 12GB during the sampling computation in the base model, and even more astonishingly, it reaches 20GB during the sampling computation in the refiner. Adding the "lowvram" option in the command line only provides a modest 1/3 improvement. This doesn't leave sufficient VRAM space for the use of other nodes in the subsequent process.

Even on webui, VRAM usage also reaches 12GB,but i thought comfyui will use less.