One last thing I would like to mention
is this: timing differences between the T&L part of the pipeline and
the rendering part of the pipeline. T&L has a reasonably predictable
timing. You know that after a specific number of cycles there will be a
new polygon ready to be rendered (drawn). Lets say that a new polygon is
ready every 10 clock cycles. Rendering itself is also predictable but its
dependent on what has to be drawn. You can draw large or small polygons.
The time needed to draw something is dependent on the size of what has to
be drawn. A small 8-pixel polygon might only take 4 cycles to render while
a 250-pixel polygon (e.g. the sky) might require up to 125 cycles! Notice
that small polygons can be drawn at a rate faster than they can be Transformed
and Lit while large polygons can be Transformed and Lit faster than they
can be drawn. This results in an unbalance between two parts of the pipeline
and an unbalance can result in one part of the pipeline waiting for the
other to finish. A Buffer can temporary help to solve this problem but its
no miracle solution. An ideal technique would alternate large and small
polygons but this is very hard to achieve in hardware. For optimal flow
a rather large buffer will be needed, these buffers use up a lot of transistors
unless the buffer is located in external ram. But locating them in external
ram eats bandwidth which again decreases performance. It's highly probably
that in most cases the render core will be the limiting factor (inefficient
memory).
In short T&L is great it can introduce a lot of extra detail in our
games but … Don't expect miracles… take all numbers with a huge grain of
salt because the numbers show will probably be marketing numbers. Don't
hesitate to divide them by a serious factor to get the real world in game
sustained rate. I personally expect that the new generation of T&L supporting
cards will approach the performance of the Sega Dreamcast console which
means sustained polygon rates between 2 and 3 million. I seriously doubt
that any product will deliver more than 5 million in true gaming situations.
An average of 1 million wouldn't surprise me at all. On top of that it will
take between 4 and 6 months for the industry to catch up with this new feature.
Adding T&L support is not a trivial job, it will require new principles
and 3D engines. T&L is definitely a step forward but it will take time
before we can see the true advantages and power of it.