the chip properly busy but not over busy and not wasting time as well".
This is really difficult, for lot of reasons ..
As an "extremely gross rule of thumb" you should consider "don't feed the
chip more than X polygons at time and more than N textures at time" and
that would very roughly work but you'd be quite surely under-use the power of the chip.
One of the biggest problems is always the same, the bus, you have those
super-ultra GPUs BUT the amount of memory they can process is still
"not that big" and the GPU-CPU transfer is always limited.
Fact is, to have "the max you can get" the GPU should be something
"integrated" inside the CPU iteslf like we moved from external Floating Point
coprocessor chips to FPUs as part of the CPU.
The day we'll have a GPU as integral part of the CPU core things could
be lot different and this seems to be the direction.
GPUs ARE really powerful but of course by the very nature of the data
they process and all it's really hard to keep them working always "at best".
Let's say so, you can have a well written program/algorithm that is all
well aligned and sorted and is making all the cache and all work as smooth
as silk and then "someone moves a camera" and *blam* everything
scrambled, all the caches invalidated, all the pipelines broken all the data
updated completely and all tits ..
That's more or less the scenario of a GPU, it's hard for various reasons
and it's all but trivial to optimize stuff to make it well.
It can be you have a fall back from 60 to 30 and then after 1 second all
is 60 again when everything is "well realigned again" but it's fucking hard
to say "this temporary drop should not happen".
Fact is, depending on what you are redering an how sometime a GPU
can well render "smegload of stuff" even at 1000 fps but .. you change
"a comma" of all that and then you see that way of rendering and feeding
up the things is not valid any more and the frame rate drops to 30 fps ..
So let's say that "30 fps should always be granted because there's such
HW grunt that the worst case scenario is still guaranteed 30" .. but
over 30 it means "you are in those special well designed cases where it
can be easy to jump out of those all in a sudden".
Despite all this power there are still a number of HW issuses well yet
to be solved, it's not perfection yet.
And we still need now the "N cores GPUs"




