> I wonder if modular memory will continue to evolve and be competitive bandwidth wise with soldered.
That's the promise of CAMM2, which is supposed to enable socketed LPDDR with almost the same performance as soldered-down LPDDR. It's still pretty bleeding-edge though so it's hard to blame Framework for sticking with soldered memory for now.
CAMM2 only has a 128-bit bus so it's going to severely compromise performance for workloads that want higher interconnect bandwidth, which Strix Halo is targeted at. For things like that, wider busses are always going to give much better performance/watt than upping clock speeds.
I'd be more than happy to see CAMM2 in general laptops, but it will probably always be much weaker at shared GPU/CPU designs like Strix Halo, Grace, Apple's M series, etc.
That's the promise of CAMM2, which is supposed to enable socketed LPDDR with almost the same performance as soldered-down LPDDR. It's still pretty bleeding-edge though so it's hard to blame Framework for sticking with soldered memory for now.