|
Below are two
photographs of an actual wafer of GeForce FX cores.
These were produced at TSMC, using their new .13u
manufacturing process. The decision to produce the
GeForce FX using this new process is the main reason this
product is yet to hit the market, but it is also one of the
reasons this part will be so powerful.
If you took the
time to read the specs that we were able to compile from
NVIDIA's technical briefs, you no doubt saw the very high
clock speeds and data rates of the GeForce FX. The
sample we played with had a default core clock speed of
500MHz, 67% higher than a RADEON 9700 Pro! Given that
both the GeForce FX and RADEON 9700 Pro have 8 pixel
pipelines, at least according to preliminary information
given to us by NVIDIA, that also means the GeForce FX's theoretical peak
pixel fillrate is 67% higher than the RADEON (4.0MP/s versus
2.4MP/s)! However, NVIDIA hasn't settled on the final clock
speed for the GeForce FX. We're also hearing through the
grapevine that core clock speeds higher than 500MHz are a
real possibility.
To attain this very high core
clock speed, NVIDIA made three other major decisions.
First, the core is built using a BGA (Ball Grid Array)
flip-chip design, with copper interconnects for better power
and ground distribution, which in turn yields cleaner signals
at higher frequencies. Second, and the most obvious,
was the decision to use a very large copper heat pipe
cooling solution, dubbed "FX Flow", similar the Abit's OTES solution. The cooling
system on the GeForce FX pulls in cool air from outside of
the case, which circulates over the GPU and is then expelled
out of the system. Something interesting to note is
that the fan on the cooler is throttled depending on the
temperature of the GPU. During light 2D use, the fan's
speed is reduced but during a heavy bout of 3D gaming, the
fan is sped up to keep the core nice and chilly. There
is a good chance every GeForce FX to hit the market will NOT
employ the FX Flow cooling system, lower clocked models, or
certain OEM derivatives may use different cooling methods,
and my not fill two slot locations in your system. Third
was the decision to add an auxiliary Molex power connector
to the card. Initial speculation was that NVIDIA
wouldn't need to add an auxiliary power connector because
the core was being produced using .13u technology. As you
can see this didn't turn out to be the case but then again,
the clock speeds are screaming on this chip and it requires
lots of power.
No matter how high the core is
clocked, if the GPU isn't fed enough data, performance
will suffer. A GPU this powerful needs a ton of
memory bandwidth to operate at or near its top speed.
NVIDIA's solution to this problem is three fold.
NVIDIA has implemented a third generation version of their
very efficient "Lightspeed" Memory Controller,
which operates with four independent 32-Bit memory
controllers, for an effective 128 Bits. In
addition, GeForce
FX boards will be populated with ultra fast DDRII-type memory
with effective clock speeds hovering around 1GHz!
The combination of this high-speed memory and NVIDIA's
memory controller offer about 16GB/s of bandwidth when
operating at 500MHz. The final boost comes from a
proprietary 4:1 lossless color compression scheme that
effectively raises theoretical max bandwidth to 48GB/s.
|
|
|
Screenshots |
|
A Vision of
Gaming's Future. |
|
The screenshots
below were all provided by NVIDIA. They are meant to
show the benefits of the GeForce FX's 128-Bit Floating Point
color capabilities and to demonstrate the increased
geometry processing power. Unfortunately, you aren't going to be
able to see much difference between these shots and ones
taken on any other high-end video card. Odds are
you're running your desktop at 16 or 32-Bit color right now,
plus these are compressed JPEGs, so any benefits realized with 128-Bit color precision are lost when
viewing these images through your browser. What they will
demonstrate though, is the higher level of detail, and
realism the "next generation" of graphics processors are
capable of.
There is no
denying that these screenshots show a level of detail an
order of magnitude greater than the games we're currently
playing. Pay special attention to the realistic
shadowing and increased polygon count. The Rallisport
Challenge and Splinter Cell shots are especially realistic
and impressive. The models are casting realistic shadows and
are comprised of so many polygons, the blockiness we've
grown accustomed to is virtually eliminated.
To further cleanup images,
NVIDIA will also be introducing new antialiasing methods
with the GeForce FX. Their new "Intellisample"
Antialiasing engine offers up to 8X AA and also benefits
from some "architectural enhancements". We hope to be
able to show you the benefit of this new engine, in the weeks ahead,
once we have a final product to test in the lab. As we
mentioned earlier, the GeForce FX also does 4:1 "lossless"
compression of color information, as well as faster clearing
of the color buffer. Like the R300, the GeForce FX
also dynamically corrects the gamma levels of the AA samples
for smoother, more accurate color values. The end
result should be some of the best antialiasing yet on our
desktops, and because available memory bandwidth is used
much more efficiently the performance penalty when enabling
AA should be much smaller than with the currently generation
of products. A Bit
More Eye Candy...
|
