I think that’s the key point here. Old (pre 3.0) Cycles would allocate one device per tile but new Cycles (formerly called Cycles-X) now allocates all rendering device resources to one tile at a time.
The only option you have is whether you have Auto Tile on or off. If it’s off, then you get ONE big tile for the whole image and if that’s too much data for your memory/GPU then bad luck. So the default is for Auto Tile to be ON with the size set to 2048. If you render an image with neither dimension larger than 2048 pixels, then again you’ll just get one tile for the whole image.
If a dimension of your image is more than 2048, then you’ll get two or more tiles starting with the lower-left corner of the image, proceeding from left to right and then bottom to top (I wish there were still a way to get it to start centered in the middle). Each tile will be rendered (but not denoised until the entire image is complete I think) using all hardware resources configured and available and no work will be done on the second tile until the first is complete.
Each tile is always rendered progressively, with or without adaptive sampling. Cycles now defaults to Adaptive Sampling at 0.01, 4096 max samples, and OIDN denoising of the final image. I like to do 0.005 and 256 max samples for quick renders.