I figured to test these myself because fast Googling claims that 512 is the best for Cycles GPU rendering, and my 512x512 did do something retarded 20mins and so I did just quit it
AFAIK optimal tile sizes depend on per-device basis.
All I know is small, power of 2 (around 32x32, 64x64) tiles are supposed to be good for CPUs. And yeah higher (like 512x512) are better for GPUs. But tests like this are very much appreciated, coz it let’s people with the same compute devices know optimal tile sizes without actually doing the hard work themselves
Think its that denoising whats changing things a bit, because 512x512 basically spitted out the whole picture, and then kept on doing nothing for 6-7mins, and spitted out again clear picture, but I dont care to test it right now