As far as I understand, a number of things make a difference:
- Socketed RAM is limited by the number of pins, DDR SO-DIMM for example has 2-3x fewer pins than your average LPDDR4X module. Having more pins helps running the chips on lower power as it was explained to me by an engineer (something to do with grounding and lower voltage, I am not really sure how it works but I trust the guy)
- Socketed RAM is a least common denominator, there will inevitably be some variance between different modules, which leaves you with less possibilities for optimizing the electrical connection
- Current LPDDR is much more sophisticated than regular DDR, the protocol is different etc. So you can't just stick LPDDR chips on a DDR SO-DIMM and expect it to work
I am sure one could design a modular system based on LPDDR, but the question is whether such system would be feasible. It would be limited to premium laptops only (as modular system would take too much space in a smartphone), it would probably need a higher pinout (which would mean a complex mounting bracket of some sort) and it would require laptop manufacturers to agree on a certain standard.
There are other considerations as well, especially if one wants to break the current performance boundaries. For example, Intel has announced that they will integrate HBM with their upcoming Xeons. How do you intend to make that socketable, that's 2048-bit memory bus. Or consider Apple's new in-house SoCs. By mounting RAM directly onto the package substrate they can potentially deliver high-bandwidth RAM in an energy efficient package without increasing the cost and the complexity of the mainboard. E.g. their upcoming hardware is widely expected to use 256-bit RAM, something that has so far been reserved for workstations. As you make the interface wider and wider, modularity becomes more and more expensive. Why don't GPU's offer modular RAM for example?
- Socketed RAM is limited by the number of pins, DDR SO-DIMM for example has 2-3x fewer pins than your average LPDDR4X module. Having more pins helps running the chips on lower power as it was explained to me by an engineer (something to do with grounding and lower voltage, I am not really sure how it works but I trust the guy)
- Socketed RAM is a least common denominator, there will inevitably be some variance between different modules, which leaves you with less possibilities for optimizing the electrical connection
- Current LPDDR is much more sophisticated than regular DDR, the protocol is different etc. So you can't just stick LPDDR chips on a DDR SO-DIMM and expect it to work
I am sure one could design a modular system based on LPDDR, but the question is whether such system would be feasible. It would be limited to premium laptops only (as modular system would take too much space in a smartphone), it would probably need a higher pinout (which would mean a complex mounting bracket of some sort) and it would require laptop manufacturers to agree on a certain standard.
There are other considerations as well, especially if one wants to break the current performance boundaries. For example, Intel has announced that they will integrate HBM with their upcoming Xeons. How do you intend to make that socketable, that's 2048-bit memory bus. Or consider Apple's new in-house SoCs. By mounting RAM directly onto the package substrate they can potentially deliver high-bandwidth RAM in an energy efficient package without increasing the cost and the complexity of the mainboard. E.g. their upcoming hardware is widely expected to use 256-bit RAM, something that has so far been reserved for workstations. As you make the interface wider and wider, modularity becomes more and more expensive. Why don't GPU's offer modular RAM for example?