I invite you to compare these two images:





So these are cross sections of the A14 and A15.
The A14 is a standard InFO-PoP package, what every A chip has looked like since the A10.
Essentially the DRAM package (one or two stacks of DRAM, probably wire-bonded, sitting on a substrate, and built by Micron or whoever) is mounted on the the Apple SoC which is connected, via InFO (so very dense bonding to a very thin RDL as opposed to a thicker substrate).
It's that pair, DRAM on top, probably about twice as thick as the SoC plus RDL below, that form the A14 package.
(The two items hanging below the A14 package are, capacitors, embedded in the moulding at the bottom of the package to make better use of that space).
The whole lot sits on the PCB (many more layers that are much less fine than the DRAM or InFO routing layers.

OK, that's great. Now the A15. It's a much smaller image (sorry, that's what the internet gives!) but the similarities are obvious. EXCEPT the grey die sitting between the two DRAM stacks.
Both A14 and A15 used LPDDR4X, so that's no help. I've no idea what (in a commodity DRAM) this could be. Two interesting question are:
- what is it??? It's not small!
- does it mean that Apple is no longer using a commodity DRAM but is negotiating with DRAM vendors to include whatever this is as part of the DRAM package.

If I were casting about wildly, the only thing I can think of is some sort of PiM (Processing in Memory) Engine that can be programmed to do various simple computations on large amounts of data without that data having to leave the DRAM module. On the one hand, this is the sort of crazy five years ahead of its time weirdness we've come to expect from Apple; on the other hand they generally tend to boast about this sort of thing, at least obliquely (eg when AMX was added).

One person who at least appears to know what he is talking about has claimed that
- Apple is completely rewriting Siri from scratch (as opposed to the current version that has just grown over time) AND
- the new version (to ship next year, driven both by SW schedule and a critical mass of HW being available to host it) will do all sorts of amazing AI magic never seen before, and can do so because Apple has devoted so much (not especially visible) HW to laying the ground work for this sort of thing. Which also suggests PIM...