In years past, Graphics Processing Units (GPU) were so much faster than main memory, that to avoid slowing GPU down, we gave the GPU its own much faster Display Memory. Since that fast Display Memory was expensive, it was a limited amount, which introduced some restrictions and drove up costs.
But now we needed to move things back and forth between special Display Memory and Main Memory, which introduced different restrictions and performance bottlenecks.
------
If we could just improve memory speed enough, there would be no need for separate Display memory. Memory speeds have been slowly but steadily improving, but not by quite enough. One way to get the extra 'bump' in memory access speed needed was to make memory Busses Wider. But ordinary printed circuit board technology did not work well with wider busses. Traces were too physically large, and were subject to signal skew and difficulties in manufacturing.
Instead, the memory and processor are tucked tightly together on one 'system on a a chip' -- a Multi-layer ceramic module with the memory and really wide busses already included. This uses technology that leap-frogs printed circuit boards and goes half way toward the technology used to 'print' the chips themselves. Now ALL memory is really fast memory.
-------
Architecturally, this innovative design means Display Memory is shared with main memory, eliminating the bottlenecks and limitations. There is no longer a 'wall' constraining the size of Display Memory. There is no need to bulk move stuff back and forth. If needed, Display Memory can grow, and if not needed, that Memory can be used for other things. Performance improves all around.
This also opens another amazing opportunity: Software supported by carefully crafted System Calls can operate on the same data in memory using CPU, GPU, AND neural engine, all at the same time.
