Unified memory: The memory management units of the GPU and CPU are coherent, and thus memory does not have to be physically moved between locations for the CPU and the GPU to do work on the same data. Apple is the only one doing this currently. This Kirin the Intel Lunar Lake both have shown no indications that this is possible yet.
Memory on package: This is what's being discussed in this article. a multi chip package that includes the SOC die and all of the RAM. A common mistake done elsewhere not here is to believe that Apple's memory, it's on the die itself, part of the SOC.
Another mistake is to confuse unified memory with shared memory. Intel an AMD SOCs with integrated graphics have shared memory, that is both the GPU and CPU using the same physical RAM. However, with shared memory the GPU is assigned a portion of that physical system RAM and the CPU memory management unit is not allowed to touch the GPU RAM. Data from must be physically copied the same way. It is copied into a discrete GPUs VRAM even though both units are using the same physical memory dies.
interestingly, an AMD GPU engineer revealed the AMDs Ryzen APUs are capable of unified memory. AMD did the work to make unified memory possible on the last three Bulldozer APUs (Kaveri, Carrizo, and Bristol Ridge.)
Apparently, that capability carried through into the Ryzen APUs, but they never implement it it into the drivers or compilers until the datacenter MI300A. All consumer Ryzen APUs function as shared memory units, unfortunately. Apparently, Microsoft is not yet interested in supporting it and AMD doesn't want to make the effort for driver implementation for Linux alone.