It’s probably the increased capacitance (think of of it as a puddle that needs filling before the water can move beyond it) of a mechanical connnection system vs direct soldering that makes most of the difference.

I was going to call you out on the distance thing but I made the maths and indeed at 100GHz light only travels about 3mm between waves and electric signal propagation on a line is roughly lightspeed (if you disregard capacitance) so even though this memory bus is likely not working at 100GHz to get 100GB/s (it’s actually using paralellism for increase bits per cycle) it is none the less already within clock speed ranges were distances of centimeters do mater.

That said keep in mind that rountrip propagation only really maters at the very biginning of the download of a memory block as that when the address goes down and the data starts coming back and the roundtrip propagation affects the delay between them.

But yeah, I can see how you would start worrying with centimeter and even millimiter distances when trying to extract a bit more performance from data exchanges at these clock speeds.

The PC GPU does have it’s own soldered RAM. But then the performance of a good GPU goes way past that of a MacBook, which while good for integrated graphics, is still only on par with a GTX 1660, a four year old budget GPU.