It still shows up on 16bit int, float, and 32bit float. But indeed 8bit is working just fine.

16bit int is less frequent than 16bit float, 32 bit float is of course the worst.

@sketchstick remarked that the problem might not be the bitdepth, but rather the profile. The linear sRGB profile gives flicking in 8bit, indicating the conversion might cause slowdown.