I think the only actual performance downsides once everything is already loaded into vram should be from the sampled values being less often concurrent in memory (which shouldn’t matter at all if mipmaps are being used)

What other step could decrease performance?