I always find it amusing when people cite the number of pages of a datasheet or other spec - be it in the hundreds or thousands - as a valid reason why no-one in their right mind would want to look at them, ever, heaven forbid. Actually, only laymen could possibly think that those documents should be read page by page in their entirety before being of any use. In fact, far from it. It's true that prolonged use of a particular chip will slowly get you acquainted with much of such a document, but datasheets / user manuals / etc. are much more like a reference - a phone book you open in a specific spot, when needing to scratch a specific itch. It doesn't matter how many hundreds or thousands of registers a chip has documented, when one usually is only directly interested in about a dozen of them at a time (unless _all of them_ have to be set up correctly before any use - that's a specific and rare case). For instance, I know my way around PIC datasheets with a blindfold on, yet I have never read the I2C multi-master part, nor do I expect to ever do it. So could we please agree to stop citing numbers of pages (or registers) as any sort of valid argument for anything?
Also, by now I think absolutely everybody save the newest of the newbies here understood what the deal with the GPU is, and (even if possibly begrudgingly) accepted it. I believe nobody is asking for that spec anymore, so "we" might consider stop being scandalized again and again about that too. What some people would still like (and I think have expected) to see at some point (="faster would be better", and yes that's really a quote) is the lowest possible (register level) spec of the allegedly non-secret, trivial GPIO parts of the SoC, which seem to have been declared anywhere between "forget it", "supported with driver, just use that", "hopefully going to be supported, if someone writes a driver" and "you'll get to reverse engineer it from the driver source" at some point in time. Of course, nobody is obliged to produce such a spec, it just seems to be the belief of some that at some point it was understood it would become available. Perhaps they were all wrong all along, it's quite possible frankly.
However, as someone coming from a hardware background to this project, I too am one of those who are accustomed to work with the silicon directly whenever we see fit, and have all specification - global architectural, part and register specific, electric, thermal, physical and everything else you can possibly think of up to and including SMD tape an reel positioning - at our disposal; a condition generally met for the vast majority of MCUs up to and including ARM class. Not being complete idiots (surprise) we do understand this is not really going to work in this specific case (and in many similar cases where GPUs are involved), but we - I at least - were definitely hoping for more than a software API and access to the sources of the drivers.
Frankly, after all the discussion on this I'm still very much confused about if and how much of a more specific spec is about to become available, but I'm not writing this trying to demand anything - I'm way past that. Just trying to clarify what I (and a few others I think) were saying / asking for all that time. I came here looking for a step-up from the standard 8/16/32-bit MCU (Arduino class if you will), looking for the same just with more oomph at less than daylight robbery prices (like a certain canine-related board et al.), and it starts to slowly dawn on me this might not be it - entirely not for architectural, but for mere legal reasons. I need hardware that I'm allowed to control. And honestly, I still have no idea how much of that I'll get to have here. That's all, thanks for reading.