1. memory where the cost for the GPU/CPU to read/write to/from it is identical (and very low)
2. no more things like "lock buffer/fill it/unlock it/render it" and/or even worse "fill this buffer by the CPU/copy it somewhere into the GPU"
3. more things "interrupts/event like" to know when the GPU is done with something and something else can be "injected" in
4. possibly be able to generate geometry on the GPU in ways that are more efficient than the CPU itself
5. possibly be able to modify geometry on the GPU in ways that are more efficient than the CPU itself
In theory all points 1 .. 5 are kinda covered on the PS4 .. once you find "the best way" to do what you want.
The fact is 'everything could be done CPU side', I could simply generate "tons of polys on the fly" and make all the bloody geoms I want and constantly pump data into the GPU.
Fact is, so far, when we speak about "mobile" as in iPad/Vita/android in some way you kinda always find out that the GPU for graphics tasks is generally faster than the CPU and CPUs are quite a bit "not so powerful" when it turns down into arithmetic and such, even just simply "have a buffer of polys, apply a simple transformation on them" is veeery consuming and it seems to be really BAD on ARM based architecture ( no FPUs ? ).
So you end up in "doing all you can GPU side" ( for the graphics ) .. and then again what I "normally" saw is :
1. pixel shader performance are generally "bad" on mobiles
2. alpha blending seems to be a mega curse on mobiles ( except Vita )
3. texture performance is quite bad as well
Overall "mobile" performance is not that great and you don't really know that well how/when to "split the damage" between CPU/GPU and quite often there are "strange surprises".
The fact is "since the pixel shader performance in term of computation" seems to be 'already monstrous' by its own, considering you can apply sin/cos/math/crap per pixel basis, would make "a lot of sense" to think "wow, if it can do that per PIXEL, should be 'piss easy' per VERTEX given that there always will less vertex than pixels".
And one thinks "if they can do that in a vertex shader" so "in theory whyt that can't do it so well and fast on a GS as well ?"
I mean the way one THINKS - without knowing too deep into - the way it works is :
mem -> fetch vertexes -> vs - > ps - > raster
Now if you'd have "a super mega fast cache on the GPU" let's say .. 64K of mem or so .. one would suppose/think that putting thing in place like :
mem -> fetches vertexes in cache 0 -> gs -> output in cache 1 -> ps -> raster
should make "almost no difference" in the sense that "ok you have one extra step in the middle" but it's like "you are fetching from one mem instead of another".
Actually "in abstract" one could even think "IF id' have TWO CPUs instead of one" I could do this :
CPU 1 -> does shit create vertexes -> mem -> CPU 2 -> Does the GS -> graphic card ..
So .. "why you can't optimise the GS as well " ?
[edit] - Actually I forgot , we should also add :
1. throw away Windows
2. throw away Android
3. throw away iOS
4. throw away Linux
5. actually throw away any "OS" at all just let me run my appl
Well on the Vita / PS4 we KINDA are there, at least 1 .. 4 are satisfied

