Great idea - I'm currently using a barebones Framework 16 mainboard with the 7940HS as my homeserver and local-AI playground in a 3D-printed case.
UMA with the iGPU is nice though not much faster than CPU (20-30%? don't recall).
But installing 64 or even 96GB RAM is more attainable though slower (60GB/s) than Apple's $$$-mem-upgraded M* Pro (200), Max (400) and Ultras (800). Although DDR5 is comparable to a regular M1 (70) afair. Most LLMs in the 20GB range run at around 7-12tok/s which is fine for playing around with them.
