/*
The test I've read said that it doesn't worth the cost for now : in fact
even AGP 4x isn't used at it's full potential, years after its release,
AGP 8x the same and even the same for PCI Express which was just
released a month ago. I have even see that you loose perf with DDR-2 !
*/
It does not come as a surprise that AGP x, 4x et cetera are NOT used to
their "full potential" as it is much more efficient to keep the working set
in GPU local memory. Initially, when AGP was launched some vendors promoted
use of "DIME" feature of AGP, including Matrox Graphics with their G200
series of products. While DIME was "okay" speed it still was a lot slower
than rendering from local memory. During the years the gap between the two
has just grown wider.. using AGP 8x to it's "full potential" just doesn't
make any sense, unless there is data constantly generated by the CPU and
pushed to the GPU.. even this is less practical now that GPU's are much more
programmable than the fixed pipeline model that we have to deal with just a
few years ago.. basicly we can do with a much smaller seed data and
synthesize data in the GPU.
Introduction of VS 3.0 with vertex samplers is important step since now we
can synthesize data with pixel program, render into texture and use that as
input to next pass vertex program. Latencies these techniques introduce can
be amortized in practise.
What PCI Express brings to the developers is a window to completely
different kind of uses of GPU the most important factor in this is that the
bus is full bandwidth to both directions, finally it is feasible to read
back from GPU memory to system memory. Thinking PCI Express only as
increased bandwidth is very narrow-minded and shows that whoever makes these
claims does have shallow understanding of modern GPU programming. Propably
some journalist without technical background, just a lucky guess..