The reference plane is the nearest metal (whether it is on the board or off it) in any given direction in 3-D space. But the effective reference plane(s) will be those that are closest to the signal trace. In the image that you give the nearest metal to layer 8 signals is of course layer 7 which is why we defined that this should be a GND plane.
As layers 2 and 4 are also both GND planes, the currents that the signals on layer 8 induce are in the GND plane on layer 7. From this it is immediately obvious that although layers 2, 4 and 7 are all called GND, and are coupled together, they definitely do not have the same currents flowing in the same parts in the same directions. They are different, but linked.
Although we say that the current flows “in” the plane, to be more precise it flows “on” the plane and the penetration depth of the current is usually very small. This means that one side of the plane and the other side of the plane will have different currents flowing! The bottom line is that the energy transfer takes place in the gap between the plane and the signal layers, and the metal on each layer is acting as a waveguide to direct the path.
Since the GND planes are clearly not the same, what happens when the signal transitions from layer 8 to layer 1? After all, both of these are GND plane! Yes, they are GND planes but the signal only references the “nearest metal in every direction” so the current you are causing to flow in the layer 7 GND now needs to flow in the layer 2 GND. How can it get there? It uses the vias to change layer. These vias should be as close as you can possibly get them to the signal via linking layers 1 and 8. Ideally the GND linking via (or better, vias) is right next to the signal via, but life is often not so kind and you may have to put it further away while realising that this is less than ideal but better than none at all.
At this point you are now very grateful that both layers 2 and 7 are GND, because coupling the planes that the displacement currents are flowing in is now trivial. One or more vias next to the signal via solve the problem. What if one of the planes was say VCC? A coupling via becomes a power to GND short. One option is link them through a capacitor. The problem is that the impedance of all decent decoupling capacitors is high at the frequencies that are being induced in the planes by your signals. (Look at an impedance plot for a typical 100nF 0402 part and the inductance of it dominates above 30 MHz or so - but what frequency are your signals - 10x, 100x?)
Let’s also briefly look at what happens for signals on layer 3. Their displacement currents flow on layers 2 and 4. Will this be equally? Only if the spacing between layer 2 to 3 is the same as that between layer 3 to 4, otherwise it is biased one way or the other.
The only layer with a problem is therefore signal layer 6 because it induces currents flowing in layers 7 and 5. Layer 5 is the problem because it is not another GND layer. Fixing this partially is possible by spacing layer 6 much closer to layer 7 than to layer 5 and so you have an engineering compromise … but it’s better than nothing.
With this in mind, what do you want in your stack-up? It is getting to look more like this:
1 - high-speed
thin prepreg
2 - GND
thin prepreg
3 - high-speed
thick core
4 - GND
thin prepreg
5 - Power
thick core
6 - low speed
thin prepreg
7 - GND
thin prepreg
8 - high speed
This is a 2+4+2 stack-up.
The vias are now:
Through vias 1-8
microvia 1-2
microvia 2-3 (or skip vias 1-3)
buried vias 3-6 (may not be necessary)
microvia 6-7 (or skip vias 6-8)
microvia 7-8
When it comes to two layers of microvias, the “safe” route is to stagger them. The denser route is to stack them on top of each other. But this is a question you should address to Sierra Circuits as to which is “better”.
So in answer to the question, can GND on layer 7 serve as a reference to signals on layer 8, the answer is Yes.
Perhaps this also goes some way to answering the other question that was raised.
