iXBT's NVIDIA GeForce FX 5800 Ultra Review [Archive]

alexsok

14-Feb-2003, 17:15

Tagrineth

14-Feb-2003, 17:38

What the hell is this (http://www.ixbt.com/video2/images/gffx/gffx-mem-termo.jpg)?!

THe_KELRaTH

14-Feb-2003, 18:52

hehehe Looks like someone had an accident with the shaving foam!

.......

Quite indepth tests but I do think if reviewers are going to use DoomIII then it should be mentioned that NV30 / 9700Pro are not being tested on an equal footing as otherwise the NV30 is half the speed.
Comparing NV30's x8 balanced v 9700Pro x16 Quality doesn't really help either, surely it should NV30's x8 balanced v 9700Pro's x8 performance.

Crusher

14-Feb-2003, 18:54

this is done for the protection of super-high-speed memory from focusings/inductions.Generally, on some information, Gddr- ii memory is very capricious in this plan/layout.By the way, focus attention on that, with what carefulness cooling precisely memory is executed.It is not enough that the radiator from the copper alloy is used, there still there is the special thermo-padding, which ensures 100- percent heat transfer.

Joe DeFuria

14-Feb-2003, 18:59

Wow...lots of stuff in there! Looking forward to the English version, although from what I can tell from the graphs/charts, I already have one minor, and one major issue with it.

Minor Issue: As expected performance with AA is directly compared at the "same number of samples." Even though FX's quality is obviously inferior at the same number of samples. (going by the screen-shots they provide). However, I can't read the commentary, so if they pointed out that fact in the commentary, it's acceptable.

Major issue: They compared performacne (in many charts) with FX aniso at 8X, vs. ATI's aniso at 16X?! And to boot...they used ATI "Quality" / trilinear mode, vs. GeForce "balanced" and "aggressive" modes? Their screenshots pretty clearly show that from a quality / LOD perspective, 8X on GeForce is on par with 8X from ATI. Then thry further handicap the Radeon by running it in trilinear mode. I think most here (based on our observations) would say that ATI's "quality mode" is inbetween nVidia's "application / trilinear" mode, and nVidia's "balanced" mode. I don't see any scores for Nvidia's "application" mode at all...driver problem? Lack of ability to enable aniso in a program? They just didn't do it?

Again, it's hard to really comment without reading the commentary, but something seriously seems wrong in Pixel Shader 2.0 land with the FX...even in FP16 mode if I read that one char correctly...

LeStoffer

14-Feb-2003, 19:12

Again, it's hard to really comment without reading the commentary, but something seriously seems wrong in Pixel Shader 2.0 land with the FX...even in FP16 mode if I read that one char correctly...

Well, there's no difference even though their are using the FX-fast 42.68 drivers. The GF FX just don't like PS 2.0 at this point of time.

I would say they still are going full FP32 regardless:

http://www.ixbt.com/video2/images/gffx/ps-2.png

http://www.ixbt.com/video2/images/gffx/ps-3.png

THe_KELRaTH

14-Feb-2003, 19:18

Interesting, so that would confirm that for DoomIII the NV30 is using 12bit Integer then unless that test is flawed.

Joe DeFuria

14-Feb-2003, 19:25

I would say they still are going full FP32 regardless:

Certainly possible...but then that would suggest that there's still an issue with FX's 2.0 shaders.

Assuming a best case(?) sceenario where FP16 performance exactly doubles FP32 performance...that would still only put FX performance "right on par" with Radeon 9700 Pro performance.

That's despite the theoretical 50% pixel rate advatage of the FX?

Interesting, so that would confirm that for DoomIII the NV30 is using 12bit Integer then unless that test is flawed.

I wouldn't say confirm...but it would make that a definite possibility. I assume these tests are in D3D, and who knows how the GL drivers differ...

OpenGL guy

15-Feb-2003, 09:16

iXBT have posted their long awaited review of the NVIDIA GeForce FX 5800 Ultra, which includes besides the usual array of benchmarks, the latest incarnation of 3DMark 2003 (including quality comparisons) and their own in-house benchmarking suit Right-Mark3D (consisting of synthetic DX9 tests).

Although the review is written in Russian (no English version yet), the charts should speak for themselves.
Looks like they messed up the multitexturing tests to me. The 9700 Pro should be right around 2600 mtexels/s with multitexturing. The 1800-1900 they got looks like the single texturing result.
http://www.ixbt.com/video2/images/gffx/pf-1.png
Compare to the 3D Mark 2001 fillrate numbers:
http://www.ixbt.com/video2/images/gffx/3dm-fillrate-multi.png

I know something is said about "pixel filling" but the numbers still don't make sense when talking about the pixel rate.

Another thing is that 1600x1200 6x AA should work fine in 3D Mark 2001 if you use application enabled AA.

Regarding AF. These two shots show that their testing is flawed:
http://www.ixbt.com/video2/images/gffx/r9700-xmas-anis8-mip.jpg
http://www.ixbt.com/video2/images/gffx/gffx-xmas-anis8-mip.jpg
It's clear that the GeForceFX isn't using true trilinear filtering (check out how short the region different mipmaps are interpolated over), but that the 9700 Pro is. Hardly fair at all.

I wish I could read what was discussed in the other aniso testing. Again, it's clear that the comparisons are not fair to the 9700 Pro.

Argue all you want about rotational invariance. The GeForce FX isn't rotationally invariant in "balanced" or "aggressive" modes plus it's don't doing true trilinear filtering.

Basic

15-Feb-2003, 13:50

Interesting. First closeup on the heatpipes as far as I know. I actually started to wonder if it realy was any heatpipes, but this closeup (http://www.ixbt.com/video2/images/gffx/gffx-cooler2.jpg) show that it realy is. But the question is: How is the condensated fluid supposed to flow up to the hot parts?

All other heatpipe solutions I've seen have made sure that the condensated fluid will flow down to the hot parts. Look at the Shuttle HSF (http://www.anandtech.com/showdoc.html?i=1764&p=13), or the Zalman Z80A-HP installation guidelines (http://www.zalman.co.kr/english/product/zm80a-hp.htm).

It would be interesting to see the GPU temperature in high stress situations with the computer in normal position vs tilted forward so the back of the case is facing upwards, or maybe with the case upside down. To avoid having the CPU heat up the GPU, it would be best to put up some kind of shield if turning the case upside down.

Care should be taken for things that dont like to be in the wrong direction. Remove CDs from drives, and move HDs so they stay in the correct position.

Anybody with a GeFX that want to try?

Joe DeFuria

15-Feb-2003, 14:06

I wish I could read what was discussed in the other aniso testing. Again, it's clear that the [anisotropic] comparisons are not fair to the 9700 Pro.

Agreed....and you didn't even mention that they tested Radeon 9700 at 16x vs. GeForce at 8X...

THe_KELRaTH

15-Feb-2003, 14:26

OpenGL Guy.. you should post your findings on the other threads and to HardOCP, Tom's Hardware etc.

It seems to be the assumption by most that Aggressive = Bilinear and Balanced = Trilinear. All the tests results would look VERY different if Performance v Balanced were used as equal comparisons.

Joe DeFuria

15-Feb-2003, 14:34

I think we discussed this before based on other synthetic and "real world" Aniso tests , and IIRC the general conclusion was this:

In terms of "levl of detail / texture sharpness", 8X on GeForce = 8X on Radeon 9700 at most aniso settings except probably the very worst quality aniso mode (aggressive) on GeForceFX.

In terms of other quality details, like trilinear and rotational invariance, we prety much ranked overall quality this way (From Best to Worst).

1. GeForce FX "Application"
2. Radeon 9700 "Quality"
3. GeForce FX "Balanced"
4. Radeon 9700 "Performance"
5. GeForce FX "Aggressive"

GeForce FX "Aggressive" was really crappy, not really in the same league due to much lower texture sharpness.

In any case, a "Fair" High quality Aniso comparison, IMO, would be to test BOTH the GeForce FX "Application" and GeForce FX "Balanced" against Radeon 9700 Quality, at the same level of anisitropy. (sp?). Look at the numbers, and look subjectively at the image quality, and make a judgement call as to which offers the best "performance and quality."

A fair "low quality" comparison would be to use both FX "balanced" and FX "aggressive" and compare it to Radeon "Performance".

THe_KELRaTH

15-Feb-2003, 14:55

When I 1st read the Anandtech review the Aggresive mode seemed almost out of place as the quailty is so bad. If I were to set a ranking it would look more like:
1. GeForce FX "Application" / ATI Quality w/ Max Texture and MipMap
2. GeForce FX "Balanced" / ATI Performance with Max Texture and Mip Map
3. Radeon 9700 "Performance Low Texture + low Mip map"
6. GeForce FX "Aggressive" World of it's own level

In terms of visual quality, the higher the AF quality the less noticable the comparitive differences are but there's no way I could compare Performance to Aggressive - which practically every review seems to do.
When comparing images using Performance v Balanced there seems to be no noticable difference.
Also, as Balanced is Bilinear then it shouldn't be compared to any Trilinear mode.

Joe DeFuria

15-Feb-2003, 15:07

I do agree that FX's "aggressive" is really in a world of it's own. ;)

Also, as Balanced is Bilinear then it shouldn't be compared to any Trilinear mode.

Actually, Balanced is NOT "simply" bilinear. However, it's clearly not true trilinear either. The mip-map boundaries are more abrupt than true trilinear, but it's not just a step change like bilinear either. Looks like a comprimise between the two.

THe_KELRaTH

15-Feb-2003, 15:36

"Actually, Balanced is NOT "simply" bilinear. However, it's clearly not true trilinear either. The mip-map boundaries are more abrupt than true trilinear, but it's not just a step change like bilinear either. Looks like a comprimise between the two"

This was why I added the ATI's High/low quality Mip Map / Texture settings into the equation, a kind of balancer.

Heres an example:
ATI x4AF - Trilinear
http://www.ixbt.com/video2/images/gffx/r9700-ss-anis4.jpg
GFFX - x4AF
http://www.ixbt.com/video2/images/gffx/gffx-ss-anis4.jpg
GF4 x4AF
http://www.ixbt.com/video2/images/gffx/ti4600-ss-anis4.jpg

Even the GF4 seems to have much more detail - notice the massive difference starting from the bottom of the image.

tEd

15-Feb-2003, 23:05

Basic

16-Feb-2003, 01:58

Does R300 do gamma correction when filtering textures?
In that case, is it possible to turn it on and off?

The reason for the questions is the differences in the "number-tunnel"-pictures.

Hyp-X

16-Feb-2003, 17:07

Does R300 do gamma correction when filtering textures?
In that case, is it possible to turn it on and off?

The reason for the questions is the differences in the "number-tunnel"-pictures.

R300 can gamma correct textures if the application requests it. (So in most cases this means no.)
This is not a transparent change so it cannot be forced in drivers. (A lot of other things would have to be changed.)

Unfortunately the gamma correction is done post-filtering AFAIK.
So it's not like with AA (altough it could be).

Hyp-X

16-Feb-2003, 17:11

Although the review is written in Russian (no English version yet), the charts should speak for themselves.

Looks like a true "4x2".

I was thinking "maybe without textures it can process 8 pixels", but appearently it can't.

Seems like nVidia calls it 8 pipelined for marketing reasons...

Dave Baumann

16-Feb-2003, 17:25

I was thinking "maybe without textures it can process 8 pixels", but appearently it can't.

Which test was that? (I've not really studied the review)

tEd

16-Feb-2003, 20:05

I was thinking "maybe without textures it can process 8 pixels", but appearently it can't.

Which test was that? (I've not really studied the review)

http://www.ixbt.com/video2/gffx-ref-p3.shtml#p6

Joe DeFuria

16-Feb-2003, 20:12

I was waiting until the English translation (to be sure we're not missing something) before bringing this up, but yes, it does appear that one of the following is true:

1) Either GeForceFX has 4 pixel pipelines....

or

2) It does have eight pixel pipelines...but each one requires 2 clocks to process a pixel.

Number 2 is the big question mark, of course. We've been mulling over "how may pixels per clock, per pipe" for quite a while now. (In all modes...integer, FP16, FP32....)

Joe DeFuria

16-Feb-2003, 20:15

In the "lets look for a silver lining" categroy:

If the FX is 4 pipes, or 8 pipes via 1 pixel per two clocks, then the FX is about 98% efficient. :P

LeStoffer

16-Feb-2003, 20:26

Looks like a true "4x2".

I was thinking "maybe without textures it can process 8 pixels", but appearently it can't.

Intriguing - I would also like to get some more input from you on this one. But maybe we should continue in this thread: GeForce FX: 8x1 or 4x2?

http://216.180.225.194/~beyond3d/forum/viewtopic.php?t=4252&start=60

Evildeus

16-Feb-2003, 20:38

According a french site, Nvidia says they have choosen to configurate the GFFX in the drivers as a 4*2. I don't know if it's possible or even true :). hope you guys find the real answer.

Dans la pratique, le GeForce FX ne dispose pas de 8 pipelines. Il n'en utilise en fait que 4 disposant chacune de 2 unités de texturing, comme le GeForce 4 Ti ! Tromperie sur la marchandise ? Pas du tout selon NVIDIA qui explique que le GeForce FX est très flexible et qu'ils l'ont configuré de cette manière dans les drivers car ils estiment que c'est la disposition la plus efficace actuellement avec les jeux multitexturing.
http://www.tt-hardware.com/article.php?sid=3858

Basic

17-Feb-2003, 02:28

Thanks Hyp-X

Ooops, stoopid question. I guess it was kinda obvious that it should be possible to turn it on and off if available at all. Specialy for textures that don't represent colors. :) Agree that it's unfortunate that it's post-filtering, but let's hope it will progress with coming chips.

But considering that gamma-correcting textures isn't something that many have cared about yet, do you think it's likely that they've explicitly turned it on for that test? The reson I wonder is that there is something strange going on at color-edges to black in the number-tunnel test. The black areas are slightly larger on the R300 than on NV30, and it might be an effect of gamma correction.

THe_KELRaTH

17-Feb-2003, 03:07

Maybe this will help...

Q.
Nvidia mentioned that they support the option of gamma correction in shaders. Is that a usable option or a marketing ploy so that the NV30 can quote that it has Gamma correction on the box.

A.
OpenGL Guy
Anyone can do gamma correction in the shaders... This won't help MSAA however.

Sireric
Yes, as OpenGL Guy writes, anyone can do gamma correction in the shader. The R300 has dedicated HW to perform degamma from textures into the shader, and on shader output, without using any shader instructions.

However, in the AA case when you "combine" your fragments into the final pixel color, you need to read in fragments that are gamma corrected, degamma them, merge them, and then gamma correct them before sending them out to the display buffer. This is called "resolving" the multi-sample buffer. Without gamma correction in the resolve pass, the merged fragments have the wrong color. Also, if you don't gamma correct at all into and out of the frame buffer, you lose LSBs in the color (gamma correction is a form of color compression, which lines up with the displayed color). Basically, you need to do it this way to get maximum output quality.

Without it, you either lose out some color LSBs, or you get AA that has weird intensity patterns

Seiko

18-Feb-2003, 21:22

I posted this against the thread referring to the English version but I guess its just as valid here?

Nice indepth breakdown. It's just a shame they didn't include R300 Aniso x8 scores as it really would have painted a more rounded picture. I don't mind if they don't go down the claimed sample route i.e. 8 compared with 8 but then the same must surely go for FSAA? If from the reviews I've read ATIs gamma correction AA is indeed better than nVidias none gamma corrected at the same number of samples then according to the above the FX should be run at a higher FSAA to match the IQ?
I appreciate its all subjective but then why not include the full range of samples in the review, 4,8,16 etc?
The numbers speak for themselves whilst both the reviewer and reader alike can then associate the relavent set of numbers to their chosen IQ pics? Surely then we'd all be in a better position to determine the card for us as some of us won't even notice a difference between some of the higher samples for both AA and AF.

Finally, I also would have liked more notes on the actual IQ during the test runs etc. I've heard several claims of driver anomolies which may or may not have a relation on speed. I guess long gone are the days where a reviewer would actually play their favourite games?

Other than that it rates as one of the better I've come across.

EDIT, Actually I did notice a small explanation on IQ although it hardly consisted of a multitude of games.
I'd also note that a tiny line denoting R300s FSAA is better quality than FXs at the end of the review doesn't really seem fitting. Again without trying to sound too fanboyish I really would have liked the reviewer to clearly state why aniso was kept at 16 for the R300 as opposed to both 8 and 16 especially if IQ was a secondary consideration for the FSAA scores?

Any one have any R300 scores with x8 AF etc running on the same rig?