There are substantial energy gains from the physics of being extremely close to the SoC (reduced capacitance).The unified memory is a different topic.
Focusing purely on performance... how much performance gain do you think it gives? In what software? Do you think there is more than a 0.5% gain in any usage scenario?
There are probably minor gains in having a cleaner signal path (easier, and less frequent, equalization).
What is missed in these discussions is a 3rd issue, namely that if the DRAM is fixed, it can be CHARACTERIZED.
Most DRAM (this is well known in academia) has the property that it is well above spec in most respects (loses charge much slower than spec so can be refreshed less often, requires less RAM to hold the charge, etc) but every RAM chip has a few pages that let down the team and only just meet spec.
IF your RAM is fixed, then you can
- characterize in the factory
- record the characteristics in a ROM associated with the memory controller
- exploit the characteristics. The most obvious of which is refresh each page at the rate it really requires, not at the rate spec says. It may also be possible, for example, to run some of the DRAM chips on lower voltages.
Does Apple do any of this? There are good reasons to believe they do. Much of the work in this field (characterizing real DRAMs, and suggesting how to exploit their features) has been done by Onur Mutlu at ETH, and Apple has worked with him (they have at least one joint patent, and at least one paper I've seen). There are also Apple patents that suggest their memory controller is consulting a look-up ROM to decide on how the exact parameters to use for handling parts of a DRAM.
I've no idea how to quantify any of this. However it is a fact that Apple seems to get DRAM throughputs that are like 98% of theoretical maximum, and Intel gets throughputs that are like 60% of theoretical maximum, so...
(I don't know where Intel is at right now. When I last looked at this around M1 time, my recollection is Intel is at ~60%, just something ridiculously bad compared to Apple. Maybe someone has current Intel numbers?)