Hem Test Grafikkort Gainward GeForce FX 5800 Ultra

Gainward GeForce FX 5800 Ultra

14 April, 2003

A couple of weeks ago we were one of the first sites with a preview of the ATI 9800 Pro 128 MB, today we have the honour to present the fighter in the other corner: GeForce FX 5800 Ultra 128 MB. The card we have gotten our hands on was made by Gainward and is named Ultra/1000 Plus Golden Sample.

A couple of weeks
ago we were one of the first sites with a preview of the ATI 9800 Pro
128
MB, today we have the honour to present the fighter in the other corner:
GeForce FX 5800 Ultra 128 MB. The card we have gotten our hands
on was made
by
Gainward
and is named Ultra/1000 Plus Golden

Sample.

The ever so late FX got a rather cold reception by the press at first,
but much has happened since then. Firstly, this is a real retail board
and secondly nVidia just released new drivers. Maybe these two factors
can
affect our present view of this often discussed product.

Before going on to the next part we want to point out that we have had the
card during about 4 hours for testing. Because of this, we did not have
enough time to run every benchmark we planned, neither did we have time
to clock the card down to "non-Ultra"-speed.
Let’s take a look at the specifications.

There has been
a lot of talk going on about the upcoming FX 5800 Ultra Golden Sample
from
Gainward. To mention some; the more quiet cooling and the included sound
card(!) have aroused a lot of interest. One thing’s for sure; Gainward
has definitely
not been saving on accessories. And finally they’ve put some effort into
a more “mature” box design, no more sci-fi space chicks.

Gainward FX PowerPack! Ultra/1000 Plus Golden Sample
Graphics chip:	NV30
Manufacturing:	0.13-micron
Number of transistors:	~125 million
GPU clock:	500 MHz
Memory clock:	1000 MHz / 15.6 GB/s
Pixel Shader:	2.0
Vertex Shader:	2.0
Pixel Pipelines/Pixel Fillrate:	4 / 2000 MP/s
Texture units(for every pipeline)/Texel Fillrate:	2 / 4000 MT/s
Ramdac:	(2) 400 MHz
Memory size:	128 MB
Memory type :	128-bit, DDR-II
I/O’s:	VGA, DVI-I, S-Video Ut
Extra hardware:	5.1-Soundcard, Firewire card, headphones, cables, etc.
Software:	Driver-CD, Demo CD for the soundcard etc.
Full version software:	Duke Nuke ‘Em Manhattan Project, WinCinema etc.
Estimated price:	~$500+

NV30

Samsung DDR-II 2.0ns

We have not
yet had any deep introduction of the NV30 architecture on NordicHardware,
but now, as the chip has been around various websites for preview we don’t
feel very compelled to spend much time looking at the techical aspects
of GeForce FX. Instead we choose to focus the whole retail product. Nevertheless,
a few
comments
on the
architecture
won’t hurt. For example that this chip supports DirectX 9.0 and that nVidia
took one step further by having partial support for Pixel Shader and Vertex
Shader 2.0 Extended. From what we’ve gathered, nVidia didn’t quite make
it
all the way, thus they chose to call it 2.0+ which isn’t a “real” DirectX
standard. Weather the extended support for DirectX will be used or not
remains is yet to be seen.

A look at the
specifications says a great deal. As the most probably knows nVidia has
chosen
to go after insane clock frequencies this time. A core at 500 Mhz gives the
card an incredible high fillrate. Certainly some complications exist, making
NV30
not really a fully fledged 8×1 architecture, but since this isn’t a first
look at the technology itself, just a usual “retail-review”,
we will spend less time on this. The fillrate is high, in many cases even
higher
than
Radeon
9800 Pro
can offer, especially if we look at Texel Fillrate.
nVidia has, as you might know, chosen to equip NV30
with a 128 bit memory bus. To compensate this nVidia uses DDR-II memories
from
Samsung clocked at 1 GHz. Together with a 128 bit bus this means that the
card has a bandwidth of approximately 16 GB/s which is somehwat less
than the challengers Radeon 9700 Pro and 9800 Pro.

Gainward FX 5800 Ultra

The back

One of the truly
nice things with this package is that along with the card you get a 5.1-soundcard.
The chip is produced by C-Media and though we didn’t manage to find the exact
model on their web-page, you can read more about the sister model here,
the specifications are should be roughly the same. The soundcard has Coaxial
In/Out, Digital
In/Out, GamePort and all other things you can expect of a fashionably soundcard.
An unexpected but welcomed accessory.

Sound and Firewire card

In the image
above you can see the two included cards. The soundcard is as mentioned
based on
C-Media, the firewire card is on the other hand based on a VIA chip. There
isn’t much to say about the quality on the firewire card; it does what
it’s
supposed to do. Aan optic cable is included with the soundcard, and a firewire
cable is also present.

Accessories

As you can see
the pack of goodies that is shipped along with this product is for the
most part the standard
stuff. However, the three things that stand out in this product are the soundcard
and the firewire card and the quite strange headphones. I’ve never seen
anything like the headphones really; you clip them to
your ear, so they’re something between the “earplugs” and headphones. The
sound is alright, but the fit isn’t really the greatest to say the least.
One of the problems is that they don’t fit close enough, so all “outside”
noise is still very audible.

For video editing,
DVD-playback etc. the application bundle WinCinema by Intervideo is included.
Intervideo’s products usually keep a high quality, especially in the aspect
of image quality.

One of the stranger
things Gainward has chosen to include in the package is the quite elderly
Duke Nuke Em Manhattan Project. Far and forgotten from a critics vision.
Along
with the soundcard you also get a CD with, among other things, a surround
tester. Of course among the CD’s there is the common driver CD as well
as some other stuff like Gainward’s own overclocking software. Other than
that you get the manuals for the graphics, sound and firewire cards.

As you can see
in the picture you get a molex y-connector. This is for the external power

supply connection the NV-30 chip requires. The card pulls out an amazing
65-70 watts so the extra push from the external power chord is surely
there because
it’s needed. A 256 MB version of the card would have required up to almost
100 watts so nVidia and the companies who build the card have taken it
easy
with those
versions, which is probably quite natural, as not even today’s high-end processors
consume such amounts of power yet.

In- and Outputs

Gainward has
chosen to equip the card with one VGA- and one DVI-output unlike from
their earlier products with dual-DVI. Maybe this is because this FX card
is a “reference” board from nVidia. The TV-out is a S-Video output, and
as usual
you get a S-Video to Composite adapter in the package.
In the picture above you also see the outlet for the cooling. You can read
about how Gainward’s own version of FlowFX is performing on the next page.

Gainward has
as usual supplied us with a complete and high quality package of extra hard-
and software. The only thing we could miss would be a good game and Composite/S-Video
In, but that is just trifles.

Accessories/Features
We look at the amount and usefulness for included accessories and integrated features.

When it comes
to accessibility this product has some huge problems. Even if many Internet
stores have had these cards on their pricing lists for about a month or
even
longer it is practically still impossible to get a GeForce FX at this moment.
The retailers who had FX boards in stock have had very limited stock of
like 5-10 cards.
The price is also a little bit too high for a product of this class. For
about $650 you can buy a Radeon 9700 and a fast processor. On the other
hand

the Radeon 9800 Pro came out in a few stores, and the price for those is
not very nice either, about $550 for the R350.
Something that argues for this high price is a extra hardware like the firewire
card, a sound card and a pair for headphones. All these three components
are
valued to about $50 – $100 which still leaves us with a card more expensive
than most other competitors.

Price/Accessibility
We compare the price to equal products and even the accessibility among retailers.

FlowFX a la Gainward

Heatsink

Now we’ve
reached a hot topic, pun intended, namely cooling. Most of you have probably
heard the horror
tales about FlowFX’s high sound level and its extreme heat generation by
now. Well, at least Gainward has succeeded a little bit better than nVidia
when
it comes to cooling. The card makes 7 to 9 decibels less noise than
nVidia’s reference design.
Despite the noise reduction the cards fan still makes a lot of sound. More
than any graphic card and CPU cooler we ever tried. Though this is only
when we use it in 3D mode. When in 2D mode the card underclocks
itself and the fan slows down. While doing that, the card is silent, so
silent that I can’t separate it from my CPU fan (~20dB), which is more
than
OK.

Profile

More cooling

As the picture
show, we see that the plastic cover above the heatsink doesn’t cover
the
sides. This will allow for the air to push against the chip and then quickly
move away. In the picture below we see the core of FlowFX’s design. Through
the plastic cover the air is let out of an opening on the back of the
computer.
However Gainward opted not to have both out and intake apparently.

Plastic fantastic

Gainward’s
FX follows nVidia’s reference design when it comes to the 12 layer PCB,
and since we haven’t
seen any alternative solutions it is hard to give an informed opinion on
that subject. As you can see below it is a very complicated PCB. That’s
the
main reason
why nVidia manufactured the PCB themselves. Eventually other manufacturers
will probably go over to their own PCB.

PCB

We mentioned before that the cooling device is different. nVidia is aware
of the problem and has therefore provided their drivers with a warmth
meter.
Our highest measurement lay on approximately 75 degrees Celsius (~167 Fahrenheit).
It seems that the high temperature was the reason for the crashes we
got during
the tests. It also resulted in that the card’s heatsink was so
hot that you could actually burn yourself on it (which I experienced to
after having to run cold water on my fingers after having tried
to remove the card
directly after turning off my computer).
Gainward helped out and provided the card with a better thermal compound
(at first it only had so-called ‘pads’, which in most cases has a pretty
poor
conductivity).
As they took apart the card they also took some pictures and sent them to
us. After this modification the card worked as it should and the temperature

was now about 55 degrees Celsius (~131 Fahrenheit). Also, the heatsinks were
now lukewarm. Gainward thinks that maybe the cooling got damaged during
its
long transportation from Asia and that hopefully not all the cards are so
sensitive. Keep in mind though that the card we used in this test was a
test
sample that has toured half the world. ‘Normal’ cards should not have these
problems.

Dissection

Backside

Naked card

Another
problem with FX 5800 Ultra is that the cooling device takes up an extra
PCI-slot.
The only manufacturer who has found a solution to this problem is Leadtek,
Gainward is with other words not alone with this complication. A lost
PCI-slot
is very hard for us to judge from an objective point of view, some people
need every one available and some people just need two or three of them.
Because
of its size it can not fit into small chassis like Shuttle’s mini-PC though.

nVidia went too far in their chase after high clock speeds, this something
that the majority probably have realized by now. Gainward has tried to make
the situation better, unfortunately
it does not seem to be enough. On of the better parts with this Gainward
FX board is its very low sound level when using it only for 2D purposes.

Design/Cooling
We takes a look at the functional and aesthetical design. The performance and sound level of the cooling is also very important.

Here are more
pictures from Gainward when they changed the thermal compound:

GPU-heatsink

RAM-heatsink front

Heatsink, back

Finally we got
two pictures of nVidia’s reference board, when comparing these with the pictures
of Gainward’s board you will notice that they are practically identical. (If
something is different it is probably just because of different PCB revisions)

nVidia FX

nVidia, Gainward
and BitTech have a collaboration to develop a more advanced cooling. Gainward
has together with other pioneers like Hercules and Sapphire chosen to develop
a watercooling solution, a nice move indeed. We have got information about
a watercooled board on the retail market soon. It will probably, unlike the
normal FX’s, be pretty overclockable. Most likely it will also have "Enhanced
Mode" with Gainward’s included software. ("Enhanced Mode" is
100% approved settings from Gainward which overclock the board)
Below is photos of this product taken at CeBIT 2003: (Our CeBIT-articles can

be found here and here)

Gainward LiquidFX

Nice

Image Quality
is a subject that has recently opened up for quite a bit of controversy.

Primarily the discussions is about nVidia’s enhancements for Futuremark’s
3dmark03. In Game Test 4, “Mother Nature”, Futuremark utilises
some Pixel Shader 2.0 effects. Part of Direct X 9.0/Pixel Shader 2.0-specifications

concerns new standards for raneg and precision etc. One of the new standards
is the increased precision. Full floating-point precision is according
to
Microsoft standards minimum 24 bits per component (Actually it’s a bit
less than that even). This would be 96 bits in total. Normally this standard
is called FP24 (Floating
Point
24).
The
second Full-precision protocol is FP32, which would be a precision of 128
bits. ATi’s R3X0-series possesses 96 bits precision while nVidia’s
NV3X-series offers two alternatives: 64 or 128 bits. nVidia

has made some commotion about the fact that ATi does not have full support
for FP32.

Microsoft’s WHQL-standards are clear about the fact that no driver
may on its own lower the precision from full (FP24/32) to partial precision
(FP16, or INT12e.g.). Now it seems that nVidia’s drivers do just
that in Futuremark’s benchmarking program 3Dmark03, thus showing a remarkable

increase in performance.
This does not come for free mind you. There is a price you must pay as
nVidia’s
optimizations has resulted in lowered image quality and lack of certain

effects.

Here is the
summary:

According
to Microsoft standards a driver is not allowed to lower the precision
if not the application itself says that it is permissible.
In
3Dmark03 we see a clear trend: those drivers which lower quality,
which go against the rules set by both Microsoft and Futuremark, also
generate much higher results.
Still
nVidia has not to this point managed to produce a WHQL-certified driver
(or a driver that render 3Dmark03 correctly) that produces as high results
as you get with e.g. the latest 43.45

Update: Please
note that nVidia has now released yet a new driver revision called 43.51
and according to initial impressions these drivers do have the great scores
in 3DMark03 while maintaining the image quality. We weill return once we
get our hands on a GeForce FX again with some additional benchmarks.

NV30 43.45

R300 6307

The images
above should give you a good understanding of the lowered quality. The
screenshot
which is taken with the latest drivers from nVidia shows a gray gloomy
sky. To the right we see the scene rendered with the Catalyst 3.2 drivers
from
ATi using a technology called overbright lightning. Those of you with "good"
English skills shouldn’t have a problem finding what’s missing
in the left image.
The lack of lightning effects is particularly obvious in the area by the
falling leaf, but you can also see some of it in the grass, here rendered

with Pixel Shade 2.0 effects. To compare with the reference image acquired
with the Reference Rasterizer from Microsoft, follow this link.
This
is a screenshot rendered with the Detonator 42.82 from nVidia which also
renders the image correctly (although with a much lower score). (Note
that the two linked shots are not produced by us.)

NV30 43.45

R300 6307

Moving on
to the next screenshot from the Pixel Shader 2.0 test in 3Dmark03, we
also
see obvious differences. While nVidia’s card renders the scene with a rather
dull and sterile surface on the piece of wood, ATi on the right shows
us how it really was meant to look. In short, nVidia makes it look synthetic
while ATi makes it look more like real wood. A more natural and less
sterile surface is just what Futuremark have aimed for when making these
effects.
The effects on the two animals differ a bit to. Among other things the
"distortion"
effect is less obvious on the image rendered with NV30.

Is is very
clear that nVidia does something that effects the quality pretty much
in
these tests. If they force FP16 or even INT12 is hard to tell, we’d suspect
INT12 at this time. But regardless, both is still against the rules of
Futuremark and Microsoft.
That the graphic card manufacturers themselves manipulate the results in
this way is not wanted. Not only because you want to see the graphics
as
the developer intended to. In this case, when we’re talking about a benchmark,
I would call this cheating.

Since we only had a limited time with this card, I could not grab any
screenshots myself rendered with Microsoft’s Reference Rasterizer (a software
renderer
which

gives the exact result as it should be) or with any drivers from nVidia
not containing these "optimizations". Though, we can reveal
that the both (reference rasterizer and older nVidia drivers) gives a result
that looks like the pictures we see on the right, not
left.
Apart from foul play in the performance race by nVidia, we find it quite
ironic for them to complain at ATi’s "limiting" 96 bits when
nVidia themselves aren’t even near that precision with their current "optimizations".
Because of the fact above we won’t use 3Dmark03 in our tests, nVidia has
made the results meaningless.
As a little footnote, we can reveal that Doom III will be rendered with
full precision. The question is if nVidia is willing to argue with John
Carmack
when this is topical? Doom III will use standard ARB-extentions and so far
we haven’t found any support for partial precision in the specifications. He
has also through his plan files indicated that NV30 is indeed much slower
when running the standardized ARB2 path. Though nVidia replied that further
optimizations of their drivers should bring them up to par, or even better,
with R300.

Here you can also see why we did not show benchmarks with the Doom 3 Alpha
(disregarding the fact that we would not have done that in either case
because of
pure respect of id Software and the fact that it is alpha software). The
results you can see from that version is nothing that speaks directly
for
how the game will act in its finished shape and therefore meaningless,
even misleading I might add as it currently does not run with the ARB2
path.

Enough about the optimised drivers, shall we then take a look at what progress,
if any, nVidia has come up with in their antialiasing?

GeForce
FX 5800 Ultra

NV30 0x (no AA)

Radeon
9700 Pro

R300 0x (no AA)

NV30 2x (2xMS)

R300 2x (2xMS)

NV30 Quincunx (2xMS+Filter)

R300 4x (4xMS)

NV30 4x (4xMS)

R300 6x (6xMS)

NV30 4xS (2xMS+2xSS)

NV30 6xS (4xMS+1.5xSS)

NV30 8x (4xMS+2xSS)

NV30 16x (4xMS+3xSS)

We give you
the screenshots so you can compare them yourself, but for the untrained
eye the differences can be hard to spot. Here is a summary:

With
2x FSAA the results are pretty much the same, although they both have
some flaws
nVidia’s
Quincunx gives a bit smoother edges compared to 2x FSAA but the tradeoff
is the overall sharpness
At
4x FSAA ATi gives us more convincing results, it even looks better
than
nVidia’s 6xS and 8x FSAA
4xS
FSAA and 16x FSAA gives the best quality from nVidia, better than both
the 6Xs and 8X
To
match the quality of ATi’s 6x FSAA the nVidia card needs 4xS or 16x
FSAA

Since nVidia’s
16x FSAA (which, by judging by the sample patterns, isn’t even true
16x AA) mode is just there to offer an extra alternative (The performance
in this mode is weak and you are limited to 1024×768 or even 800×600)
we have chosen to use nVidia’s 4xS vs. ATi’s 6x FSAA in our performance

test with maximal image quality. We have also chosen to compare nVidia’s
4x FSAA with ATi’s 4x, despite ATi’s advantage in image quality, so
please consider that when viewing the results for 4x AA. Let’s
take a look at how the performance is affected in the different AA

modes. We’ll start off with a Direct3D game, UT2003: (Please note that
all the FSAA tests in this review is performed using nVidia’s "Maximum
Quality" setting, if nothing else is noted.)

We
have chosen the resolution 1280×1024 to stress the card as much as possible
without concentrating
too much on fillrate performance. The UT2003 test shows that you can play with
reasonable framerates in all modes up to 4x FSAA. When the combined MSAA/SSAA
modes
kick
in we se a performance drop. UT2003 showed to be too memory consuming for 16x
FSAA in resolutions above 800×600. The 8x and 16x being almost unplayable
was
expected but the fact that not even 50 fps were reached in 4xS is a bit disappointing.

Quake 3 is our Open GL test
and it also represents an older generation of games:

We see here
that 16x FSAA does its job but, as we suspected, the performance is just way
too low. Not even in 1024×768 in an old game like Quake 3 we get even remotely
close to an acceptable frame rate. 8x AA, on the other hand, works pretty well
with its 92 fps. Before now, 250 fps in 4x FSAA could only be achieved with
a Radeon 9×00 card. Apparently there is a new kid in town, although the image
quality is lowered a notch or two.

A
couple interesting things to look closer at is that nVidia has barely even
touched
their implement of FSAA since they released GeForce3. The antialiasing
that GeForce FX is using is therefore about 2 years old which within computer

cohesion is very much. Because of this FX contains FSAA with a lot of weaknesses.
One of them is that they haven’t got support for rotated grid in other
than 2x and Quincunx. nVidia has no support for gamma corrected FSAA
either. Though, the real big limit is that nVidia has not got support for
more than
4x Multi Sampling. If they want to achieve higher FSAA-levels they must therefore
begin to mix MultiSampling with SuperSampling. You can easily se this happen

because SuperSampling messes up the text up in the left corner on the pictures.
At 16x FSAA we can see that the text is totally messed up, probably due to
a bit "screwed up" sample pattern. For example, one
letter has totally disappeared.
Why nVidia has even cared for having support for 6xS and 16x in their drivers
remains to be a mystery to us: 6xS is worse looking than 4xS and still it eats
much more performance while 16x
is simply too slow to be used in any realistic situation.

While we are at
the subject, we can tell that the drivers do not support 4xS or 6xS FSAA
in
Open GL. While in Open GL, we have then chosen to compare nVidia’s 8x FSAA
against ATi’s 6x when we test maximum image quality. Since 8x FSAA is not
working in
1600×1200, these test results has been ignored. The 16x mode in Open GL also
differs in that it supports true 16x AA (4xMS+4xSS) instead of just 12x.

Antialiasing
is not the only way to increase image quality. Anisotropic filtering is
just as important in order to push the card into producing eye-candy to
the maximum. We know that the Geforce4 Ti-series couldn’t perform very well
when using Aniso, but there was nothing wrong with the image quality. In
order to fix the problems they were facing, nVidia has mainly targeted on
optimising their drivers and off course the performance gets a significant
boost thanks to the high fillrate of the FX-card. But exactly how good is
the image quality?
We start by taking a look at the trilinear filtering. This filtering is
of great importance since it is the standard filtering that all "resonably
new" games
employ if you don’t adjust it manually.

NV30 Max Performance

NV30 Application

NV30 Max Quality (Default)

R300 Balanced (Default)

Here we can
see that with the card set to default it can’t present the same image
quality as the Radeon. Since our GeForce4 Ti 4600 is out for lunch we didn’t
have time to take a screenshot to compare it with the other two. Though
we do know that it is equivalent to the FX-card when in Application mode.

Application mode is much worse performance wise compared to the Maximum
quality mode. In short, nVidia has sacrificed quality in order to increase
performance. In all honesty this is hard to spot if you do not know
what to look for but when push comes to shove it is clear that nVidia is
taking a few steps back here.
It is pretty obvious that we will have to use both modes in our benchmarks
and judging from these pictures we might not use the Maximum performance
mode at all. The question now is how far apart the different modes are in
performance.

The
differences in performance when only using trilinear filtering is greater
than expected. Apparently there is something going on behind the curtains.
One test is not enough to come to any conclusions, so lets have a look at
Quake 3 as well. Here we choose to use the screen resolution 1600×1200 mainly
to get rid of the cpu-limitations, but also to stress the card’s fillrate.

Also
in Quake 3 the difference is big, almost too big considering the differences

in image quality. But the really interesting thing here is anisotropic
filtering. As you saw above you can choose three types of texture quality.
They are
called "Application", "Maximum Quality" and "Maximum
Performance". The names are quite misleading because in this case "Application"
is the one with "Maximum Quality". To call "Maximum Quality"
maximum quality is rather strange since it obviously does not have the
best quality. A theory about why nVidia has chosen these names is that
reviewers
(like us) are supposed to compare nVidia’s "Maximum Performance"
with ATi’s "Performance" and nVidia’s"Maximum Quality"
with ATi’s "Quality".
Let’s begin to compare the Aniso quality when looking at an even surface.

GeForce
FX 5800 Ultra (8x Aniso)

NV30 Application

Radeon
9700 Pro (16x Aniso)

R300 Quality

NV30 Max Quality (Default)

R300 Performance

NV30 Max Performance

Just like
we noticed before, it’s ”Application” that looks best.
Both Max Quality and Max Performance start to mess with the mip map levels.
nVidia has chosen Maximum Quality as Default. So if you go out and buy
a GeForce FX card today and don’t change any settings it’s
this setting that is activated. Something we found a bit
peculiar since the FX is supposed to have such fantastic image quality.
The answer to this mystery is however simple: Performance. With the "Application"
setting activated the performance of the FX is pretty lame.
In this test we can see that the FX’ 8x Application and ATi’s 16x Quality
settings are quite similar. ATi’s Performance is harder to distinguish.

Although this setting pushes back limit on the mip map levels it only
works with bilinear filtering. This makes it hard to make an objective
evaluation. However, it
is our opinion that ATi’s Performance setting seems to be way better
than nVidia’s Max Performance setting.

But, all games
aren’t just about walking around a flat, empty surface. Therefore we took a
look at how the texture filtering looks when faced with a tunnel. This is a
good situation to see how the texture filtering reacts to different angles of
the textures.

GeForce
FX 5800 Ultra (8x Aniso)

NV30 Application

Radeon
9700 Pro (16x Aniso)

R300 Quality

NV30 Max Quality (Default)

R300 Performance

NV30 Max Performance

The pictures above
show how both of the cards take shortcuts in the aspect of filtering of angled
textures. nVidia’s Application setting by far has the best picture quality.
Once again it is hard to make an evaluation of the other settings available.
nVidia’s Max Performance is however down under. Therefore our tests will compare
nVidia’s Application setting vs. ATi’s Quality and nVidia’s Max Quality
vs. ATi’s Performance setting.
Before we go on let’s take a quick look at the Aniso performances as well:

Ugh.
That really doesn’t look too good for FX just then. Unbearable and unplayable.
So let’s take another look at Quake 3 which isn’t just as demanding
as last years bestseller.

While we are looking
at picture quality anyway, we decided to give you an idea of how the quality
has change since GeForce4 Ti.

GeForce
FX 5800 Ultra (8x Aniso)

NV30 Max Quality (Default)

GeForce4
Ti4600 (8x
Aniso)

NV25 (Default)

NV30 Max Quality (Default)

NV25 (Default)

Here you can clearly
see how heavily nVidia has reduced the texture filtering quality. Pretty sad
if you ask us. When it comes to image quality we also consider 2D, TV-out and
3D quality in general. GeForce FX gives more acceptable results here. As far
as we know, the TV-out chip is the same as in the NV28/NV18 series. This means
that the great improvement in quality/functionality is not in this area, though
the TV-out feature does its job. Unfortunately, since our time for this review
was very limited, we could only do a few observations here.

In 2D quality
the card is a bit better than its predecessir, but ATi and Matrox keeps the
lead that
they’ve had for some years now. nVidia is said to have improved its “Digital
Vibrance” since last time, but no big differences were found in its
functionality. The problem with colours being oversaturated instead of having
a better luster
remains and the card lacks the warm colours you get with a card from the ATi
Radeon series. Please note that these differences are small and should not
be
exaggerated. The last part of the image quality test will be general 3D quality.
Here we get pretty impressed, especially in “Application” mode.
The picture looks very “crisp” without any texture aliasing
at all. The problem is though, that when in “Application” mode
the performance drops noticeably. Not that it really matters here since
the image quality is
what counts.

Ok, let’s
look at what we got:

nVidia still
uses pretty much the same old AA/AF technique as the GeForce3 introduced
two years ago.
Good quality
in 2D, TV-out and general 3D
nVidia has lowered
the standard quality of their texture filtering since the release of the GeForce4
Ti
ATi is a big
leap ahead when it comes to FSAA-quality while slightly behind in Aniso
(when in ”Application” mode)
So called ”optimizations”
in nVidia’s drivers lowers the image quality in 3DMark03

When we test the
GeForce FX 5800 Ultra in “Application” mode (to get quality
that resembles the Ti4600 results as much as possible) we get the impression
that
FX really is nothing but an overclocked GeForce4 in some aspects. The problem
is that the “Application” mode is more than meets the eye,
and even with a GeForce4 you’ll see a great performance loss. In short,
nVidia has created
a jungle of settings and it’s nearly impossible to compare the FX with
other cards under the same conditions. A nightmare for us reviewers.

Image quality
The image quality in 2D as well as in 3D are measured. In 3D the quality of FSAA and Aniso is important. The quality of the TV-out and any other in and outputs are compared.

The image quality
tests are important since they help us understand what settings and tests to
use. What we have come up with is what FSAA/Aniso modes to compare each other
to, and that we won’t use 3DMark03 in this review. On the next page you’ll find
our test methods and our test system.

In today’s tests,
the FX 5800 Ultra has two rivals: Radeon 9700 Pro and Radeon 9800 Pro. We
also used results from a Ti4600 as a reference.

Test system
Hardware
CPU:	AMD Athlon XP 2600+ (333) Mhz
Mainboard:	Soltek SL-75FRN-L (nForce2)
RAM:	768 MB DDR333 @ 2-5-2-2 Timings: 2x 256 MB Corsair TWINX512-3200LL DDR-SDRAM 256 MB PC3200 OCZ PC3200 Rev2 DDR-SDRAM
Graphic card:	Gainward GeForce4 Ti4600 (128 MB, 300/650) Gigabyte Radeon 9700 Pro (128 MB, 325/620) ATi Radeon 9800 Pro (128 MB, 375/700) Gainward GeForce FX 5800 Ultra (128 MB, 500/1000 MHz)
Harddrive:	80 GB Western Digital Caviar 7200 RPM Special Edition (8 MB cache)
Soundcard:	Creative Soundblaster Audigy 2 Platinum
NIC:	D-Link DFE-530TX 10/100
Software
Operative system:	Windows XP Professional (Service Pack 1)
Graphic card drivers:	nVidia: Detonator 40 43.45 ATi: Catalyst 3.2 6307
Other drivers:	nVidia UDA Chipset Drivers v2.03
Benchmarks:	3DMark2001 SE b330 Aquamark 2.3 XPBench 1.03 Quake 3 1.32 UT2003 v2199

Accessories/Features:	9.5
Price/Availability:	4.5
Design/Cooling:	3.5
Image quality:	6.5
Performance:	7.5
Overclocking:	– *
Final Mark :	*6.3*

Remnant 2 får nytt, gratis spelläge

Space Marine 2 har redan sålt över två miljoner exemplar

Playstation 5 Pro – här är allt du behöver veta

AMD vill ändra om i sina arkitekturer för enklare utveckling

PS5 Pro förväntas avtäckas idag

Test: Philips Evnia 49M2C8900 – tungdriven men imponerande välvd 32:9-skärm

Test: Logitech PRO X Superlight 2 – underväldigande uppgradering

Test: Samsung 990 Pro 4 TB – nya V-NAND ger nya…

Test: Lenovo Yoga Pro 9i – RTX 4070, för kreatörer!

TechBubbel 185 – Sista TechBubbel

TechBubbel 184 – Nya från Qualcomm: snabbare än M2

TechBubbel 183 – Intels löjliga lansering

TechBubbel 182 – Nya PlayStation 5 är ”Slim”

TechBubbel 181 – Ryssland-mobilen, i Sverige

Gainward GeForce FX 5800 Ultra

Nyheter

Remnant 2 får nytt, gratis spelläge

Space Marine 2 har redan sålt över två miljoner exemplar

Playstation 5 Pro – här är allt du behöver veta

Performance
The performance we grade compared to products in the same price class, and those close to it. For mid range products and up, performance in the quality tests carry weight.

Overclocking
Here we see how much we can overclock the card without suffering from artefacts, and then measure the performance against the default results.