You're still looking at these GPUs as though they are designed with the same kind of architecture (and are essentially the same kind of GPUs) that are in current Intel Macs and x86-64 PCs. They're not. Apple is throwing the GPU playbook out of the window with its SoC GPUs. Plus, they've stated that the RAM being used will be shared with the system memory (which I presume will be the case on at least all non-Mac Pro Macs, if not ALL Macs). This is why I'm confused about why we're talking about an AMD VRAM technology when AMD GPUs and Apple Silicon integrated GPUs are about as Apples and Oranges as you can get.
I'm struggling a bit to put this in a way that is guaranteed to be inoffensive, but the above signals serious confusion.
First off, Apple is not "throwing the GPU playbook out the window", the most notable difference with their GPUs vs. NVidia/AMD products is that Apples GPU:s are Tile Based Deferred Renderers (TBDR), as opposed to Immediate Mode Renderers(IMR). This is by no means new, TBDRs has been around over three decades, and many seminal patents are in the public domain.
Cutting it short, TBDRs can save some memory traffic for certain operations, which is nice for SoCs, but the difference vs. IMRs are not night and day, and TBDRs come with their own caveats. You'll notice that even though TBDRs have been around since the beginning of dedicated graphics silicon, they haven't taken the world by storm. Apple has a good
implementation, but that's it. And yes, they still need bandwidth. I you want to dive deeper, Apple has published a bunch of videos, not least this WWDC 2020, but ImgTech has good stuff as well. They are both quite biased as to the advantages obviously.
There is no such thing as VRAM technology. It doesn't exist. In the days of yore (80s) VRAM was shorthand for dual-ported RAM, but the concept died long ago. Fundamentally DRAM is DRAM, the differences lie in packaging and communication protocols. For PC systems with "integrated" graphics, the GPU and CPU share system RAM which unfortunately is the same as the CPU would otherwise use on its own causing serious congestion. Today, that typically is dual 64-bit channels of DDR4, which seriously hampers performance, but allows easy expansion via DIMMs and really low cost.
GPUs really need higher bandwidth RAM systems though, so when Sony And Microsoft designed their new gaming consoles, they supply their SoCs with just under 500GB/s (which is still a significant drop in bandwidth/FLOP compared to the previous generation), in order not to hamstring their performance too much. That necessitates GDDR6 that has higher frequency signalling, and signal integrity makes putting the DRAM on the PCB rather than in DIMMs mandatory, and you loose any post sales upgradeability. Which is par for the course in graphics, of course. Increasing bandwidth further is difficult with this approach, as you have to go wider meaning more complex PCBs, more area dedicated to driving I/O pins on the GPU and significant power draw from the memory subsystem alone. Still, it is done for high end offerings.
A better solution is to put the memory even closer to the SoC, using an interposer, avoiding the need to drive signals through a PCB. This allows your data paths to be wider, and thus signalling frequencies and power draw to be lower. There is additional expense, but also cost savings in PCB design and power supplies.
Again, there is nothing that says that these higher bandwidth solutions cannot be used for SoCs as well as GPUs, in fact it is already done! The major issue is that the PC infrastructure isn't built around such a paradigm, and users cannot upgrade memory capacities post purchase. Also, absolute maiximum RAM capacity is constrained. But I would trade the capability of having 256GB of RAM for an order of magnitude or more in memory bandwidth any day of the week and twice on Sundays, both privately and professionally. And if Apple wants to be able to compete with discrete GPUs, whether in 3D rendering and general compute, it is an absolute necessity.
As you say, the RAM will probably be clocked a bit lower in a laptop, more like 2-3 watts per die. Pair it with a 2048 bit memory interface and you still have some incredible memory bandwidth in a compact laptop. Imagine what having 300GB/s memory bandwidth will do for CPU performance.
