The grandparent used the wrong term. The memory isn't on the die, but it is in the package (MCM). I don't think that provides faster throughput, but it should help latency.

Regardless, the big win was giving unified access to that memory using a memory fabric. That lets the GPU (for instance) access graphics assets with no copying - it just uses them in place.

AMD can absolutely do the same thing, if it desires. Intel can too, but it has a hill to climb WRT GPU performance.

Apple's AI chiplet is another matter...