My GeForce FX commentary - a different view indeed

[Uttar]

Hey,

I'd like to comment on the GFFX. But then, why does this deserve a new thread? Well, that's quite simply because I'm going to give a completely different reasoning, and conclusion, than everyone else :) And no, I'm not going to say the GFFX is the greatest thing ever. It isn't. So read on...

First of all, let me say Anand did excellent IQ tests. IMO, he just didn't interpret the performance / IQ correctly.

Let me begin by talking about Antialiasing. Let's see what the respective performance hits are, shall we?
http://www.anandtech.com/video/showdoc.html?i=1779&p=13

GFFX 4x AA: 24% performance hit
Radeon 9700 Pro 4x AA: 25% performance hit

As you can see, the GFFX AA is *cheaper* than the Radeon 9700 Pro. Why? After all, it got 20% less bandwidth...
The answer, sadly, isn't so pretty. The GFFX 4x AA patterns are ordered, giving lower IQ and more repeating colors. Notice the "more repeating color" part? That means the SAME compression algorithm would work better on the GFFX 4x than on the Radeon 9700 Pro 4x!
However, in most cases that shouldn't make much of a difference. So the GFFX performance hit with the Radeon 9700 Pro pattern might be more like 27% or something.

Now, why is 6xS so darn slow? The answer lies the the SSAA part of it. Color Compression is a LOT less efficient with that, because not all subpixels are identical. The Radeon 9700 Pro 6x AA got all subpixels identical, so it's really normal the performance is a lot better.

4x AA scores for the GFFX, in fact, are amazingly good seeing how the GFFX got 20% less memory bandwidth than the Radeon 9700 Pro. Very few reviews got 4x AA only scores, but two do. Hexus & Tom. Both sites conclusion is ridiculous and the AF comparaison is not fair at all ( ATI got a signifiant disadvantage ) , but I doubt the benchmarks for AA are tricked.
http://www.hexus.co.uk/review.php?review=497&page=8
http://www.tomshardware.com/graphic/20030127/geforce_fx-19.html#4x_fsaa
http://www.hexus.co.uk/review.php?review=497&page=6

As you can see, in the worst case the GFFX got a 11% lower score than the Radeon 9700 Pro, with 20% slower memory. On average, they're actually *on par* with only 4x AA!
Sounds like ATI is getting a little bit more waste than nVidia with their 256-bit memory bus.

Conclusion for AA: The GFFX got two big disadvantages. Worse 4x AA patterns, and no 6x true MSAA mode. The "no true 6x MSAA mode" is only a problem if you run at lower resolutions, because seeing the difference between 1600x1200 4x AA & 6x AA is just impossible IMO.
As for the worse 4x AA patterns. That's what 4xS is supposed to fix. But the problem is that 4xS also got SSAA in it, so performance will be a lot worse. That, thus, is a huge problem.
I already said the 2x AA quality is a bug in another thread, and that it should be fixed easily with drivers. So I won't count that as a disadvantage ( it'll be fixed when we'll get to buy the card )

Maybe 4x AA is really rotated, who knows. nVidia seems to claim that AA&AF things are happening after the frame buffer, so maybe that isn't captured correctly either. But until I get a proof of it, I refuse to trust that.

Aniso now... Let's look at the numbers.
http://www.anandtech.com/video/showdoc.html?i=1779&p=10

Now, we can see nVidia's biggest problem. The HUGE performance hit when using Balanced, and the bad IQ when using aggressive. The solution, for nVidia, would be to make an intermediary setting between Aggressive and Balanced, and I really hope they'll do so. However, I fear that's unlikely :(

Right now, there's really two things thatc can be done to compare Aniso. In case you want the very best quality, you compare nVidia's 8x Balanced with ATI's 8x Performance. There, the GFFX loses very badly.
The other solution is to compare nVidia's 8x Aggressive with ATI's 4x Performance: this gives significantly lower IQ, but the IQ is the same on both cards and the performance hit is also nearly identical.

Conclusion on Aniso: Right now, nVidia got a clear disadvantage here. It is the GFFX biggest problem. I urge nVidia to implement an intermediary mode between Aggressive and Balanced ( preferably via drivers, but if it's not possible that's essantial for the NV35 ) - otherwise, they've badly lost that battle.


Now, there are several other things people have complained about, or said which were quite anormal. What are those, you ask?
First of all, some of 3DMark's 2001 SE synthetic scores are *not* what most people would have supposed based on paper specs. So, what happened?
http://www.tomshardware.com/graphic/20030127/geforce_fx-27.html
The Radeon 9700 Pro actually scores a *higher* Single Texturing score! Why? Bandwidth.
With my GF4 Ti4200 at 250/250, my single-texturing score is of 861.8MP/s
At 250/200, my score drops to 726.6MP/s due to bandwidth limitations.
So, that one is explained and drivers won't be able to do much about it: 3DMark 2001 SE fillrate tests are ridiculous and false.

Now, there's also the Advanced Pixel Shader test which uses PS1.4. where ATI got a nice lead. I've got no data to back this reasoning, but anyway...
It was revealed that DX9 was not working in those drivers. That mean PS1.4. support is probably pretty early in them, and not much work was done on it yet. So it sounds reasonable to say that nVidia has done few optimizations, and it's the area they've focused the less in.
Thus, improvements will mostly be seen in such areas.

As for a slower VS speed than the Radeon 9700 Pro when using complex shading programs. It's hard to say, really. Those programs aren't *so* complex, and bandwidth might be an issue. Also, the GFFX shading architecture is really optimized for very long programs, so I really think those programs might not put it in its best light.
Of course, huge programs won't be seen in games for a while. So those synthetic benchies most certainly show us that the GFFX VS is a little bit inferior in games which are going to release soon.


Finally... The dustbuster, aka Flow FX. Well, the noise really seems bad. But I guess that was known since they announced the whole system. However, I'd like to see lower 2D clock rates. 300/300 is completely useless in 2D! 250/250 is really a maximum, and it would generate less noise. Okay, so you can change that yourself with coolbits. But not everyone is going to do that, and I think it would be a good change.
Also, it would be nice if Flow FX could automatically put the card to something more like 400/400 if it sees FPS is more than, for example, 100. That obviously can't be done via drivers, but it's already on my NV35 wishlist.

Conclusion: On the feature side of things, nVidia is a clear winner. The GFFX is future oriented, and got a lot of cool things for developers. It's unlikely, however, that they will be used in the very short life of the card.

AA Performance is on par with what I expected. AA IQ, however, isn't: I had hoped for a pure MSAA mode with a rotated grid. 4xS performance hit is probably just gonna be too big.
AF Performance is good too, but AF IQ is not. That means AF is really a lot worse than what I had expected.
nVidia got a fast algorithm with the 8x Aggressive mode, but its IQ is on par with ATI's 4x Performance mode and that's the best the adaptive modes can do! Thus, many reviewers choose to use "balanced" and that's why the GFFX is losing so much in all those benchmarks.

The dustbuster is a problem, but it's not huge.

For the GFFX to satisfy what I hoped for, the following should be done ( in increasing order of importance ) :
- Decrease default Flow FX 2D clocks to 250/250 instead of 300/300
- Fix all reported bugs. There seems to be a fair bit of them.
- 5% to 10% performance increase in no AA/AF performance.
- 4x AA mode with pure MSAA and rotated grid sampling. 4xS performance is probably going to be too low.
- AF mode which is a compromise between Aggressive and Balanced. Aggressive is too aggressive, and Balanced performance drop is too huge. We need something between both!
- Send me a NV35 prototype before every other tech site :D j/k

Right now, I'd give the GFFX a mere 65%. If those things were done, I would easily give it 80%, or even 85% if the performance increase they can do via drivers is more like 25%. That is, if I was a reviewer :p
No matter what, I'll probably buy a GFFX if at least the three first things are done. But I might decide not to take the Ultra, not sure if it's really worth the extra money.

Anyway, that's my opinion. As you can see, it's kinda different from most other people's opinion.

I hope you liked reading this. Yes, it's long, and maybe I could have made it shorter. But I really felt like saying all of this, I've been thinking a lot about it. Anyway, thanks for reading!


Uttar

[Hellbinder]

Uttar, an intersting view.. however i would like to point out


As you can see, in the worst case the GFFX got a 11% lower score than the Radeon 9700 Pro, with 20% slower memory. On average, they're actually *on par* with only 4x AA!
Sounds like ATI is getting a little bit more waste than nVidia with their 256-bit memory bus.


Your conclusion is faulty. You already pointed out one reason for the lead in your own post. Nvidias orderd grid uses a lot of repeating colors, that results in better compression, but also results in lower IQ. Its not the memory buss usage at all.

Another point that no one except those with a keen eye will notice is that Nvidia is currently not even using a true 4x pattern, nor a 2x pattern. Thus further diminishing their iq, and allowing for slightly fewer samples, thus resulting in even less bandwidth use.


The dustbuster is a problem, but it's not huge.


You are only kidding yourself here. That fan is going to drive 90% of the people batty.


The Radeon 9700 Pro actually scores a *higher* Single Texturing score! Why? Bandwidth

I have a totally different Theory for you on this.

3dMark Fillrate test multitexture

From extremetech.com

GFFX ULTRA 3557.8
9700 Pro 2564.1

From Anand

GFFX ULTRA 3477.9
Radeon 9700 2536.7

Combine those results with the single texture results. The 9700 Pro is clocked at 325 mhz engine. That gives a theoretical maximum of 2600 in the 3D Mark multitexture fillrate test. The actual score is 2564, which is 98.6% of the maximum. Anandtecchs scores are nearly the same. Now look at the GeForce FX. Theoretical maximum at 500 mhz would be 4000, yet it achieves 3558 which is about 89% of the maximum. That seems very odd. Now, take 3558 and round it up to 3600 and now compute the clock speed: 450 mhz. Quite frankly, the multitexture test should be a very reliable way to calculate clock speed.

What conclusion do i draw from this??

By the time that the GFFX gets to the fillrate tests, it is already overheating to the point that the clock speed is getting droped by an automatic 50mhz. But this is just a theory you know. I could be wrong.

If true, it bodes very bad things for those that plan on actually running games with the GFFX.

[-=DVS=-]

Well Uttar i am sure Nvidia will send you GFFX after readng this :p

But in my Opinnion , after waiting so long and getting such a poor card is very dissapointing , NV30 already running at its maximum core with no headroom for overcloaking , unless offcourse they change core use low-k d.. on .13

And r300 at alot lower clock speed on .15 with less noise, easly stays in same speed range :eek:

And becouse Geforce Fx is new card from Nvidia drivers are with bugs , not as good and mature as GF4 or Radeon 9700 PRo
example: texture problem in some game mentioned in reviews and inferior AA/AF quality *hopefully for Nvidias sake all this will be fixed by time GFFX willl ship to stores*

Right now it stand
Radeon 9700pro vs Geforce Fx

Both in about same performance range except
one runs very hot and noisy while other is quiet and cool
One have early drivers other matures
One cost alot more then other
*something to think about*


sorry if someone thinks i am trolleing , but all this is a fact, try prove me wrong ;)

[Uttar]

Thanks for the feedback, Hellbinder.

About 4x AA: Well, less color variations only do a very slight difference in performance. About 80% of the scene is not edges I'd guess. And it's only in about 50% of edges that it might be aliased with rotated and it isn't with ordered. So that means 90% of the color compression is done with maximum efficiency.
And most of the AA performance cost remain Z Compression. So, I'd really say this only does a very small difference, and even with pure MSAA rotated, you'd get only a 14% performance drop ( compared to 11% ) , and that's a pessimistic view of it.


About the dustbuster: Well, it isn't a huge problem to me. I do agree it could be very annoying for some people. For me, however, I don't care too much about noise when playing ( unless it's a RTS, and then I could always underclock the card to get less noise )
What's important to me is to be able to concentrate when I'm not playing. That's why I'm suggesting the 250/250 clock speed when in 2D.

About fillrate: Not sure of this, I might have to check it again, but didn't Anand say that the GFFX couldn't decrease clockrate in small increments? And that if it was overheating, it would automatically go back to 2D speed, so 300/300?


DVS: I agree the GFFX isn't very attractive right now. That's why I'm giving it a 65% in the current state, and giving it a conditional much higher score if nVidia fixes multiple things.


Uttar

[Hellbinder]

About 4x AA: Well, less color variations only do a very slight difference in performance. About 80% of the scene is not edges I'd guess. And it's only in about 50% of edges that it might be aliased with rotated and it isn't with ordered. So that means 90% of the color compression is done with maximum efficiency.
And most of the AA performance cost remain Z Compression. So, I'd really say this only does a very small difference, and even with pure MSAA rotated, you'd get only a 14% performance drop ( compared to 11% ) , and that's a pessimistic view of it.


The % of the screen has little to do with the point being made. At a perfect *one* color sample, Frame-Buffer compression can compress up to 24:1 per ATi's listed spec, i imagine that this is the same for Nvidia. Thats where ordered grid is getting its big advantage for Nvidia. Which again is being assisted with, i believe a lesser sample pattern than perscribed. Again a the cost of IQ.

The total variation you are talking about in the first place is only 9% Right? Which is perfectly understandable. Considering the advantage Nvidia has due to its method.

[Uttar]

Originally posted by Hellbinder
The % of the screen has little to do with the point being made. At a perfect *one* color sample, Frame-Buffer compression can compress up to 24:1 per ATi's listed spec, i imagine that this is the same for Nvidia. Thats where ordered grid is getting its big advantage for Nvidia. Which again is being assisted with, i believe a lesser sample pattern than perscribed. Again a the cost of IQ.

The total variation you are talking about in the first place is only 9% Right? Which is perfectly understandable. Considering the advantage Nvidia has due to its method.

I don't quite think you understood what I meant... Let me retry.

On 80% of the scene, the sampling pattern will make NO difference because it's fully inside a triangle. Color compression will always be optimal in those cases.
And in many cases, ordered will give not 100% identical subpixels for all of those edges. So it doesn't make all that much of a difference.

As I said, the waste ATI is getting with their 256-bit bus is minimal. It's also very impressive. I supposed a 10% or 15% waste originally, and it sounds like it's more like 5%


Uttar

[YeuEmMaiMai]

In order to fairly compare the efficiency of the R300 vs NV30

1. clock them the same
2. make sure memory bandwidth is the same


The reason that NV30 pulls a head is IT HAS A HIGER CLOCK RATE to perform calculations at

NV30 is inferior and it's design shows this

triagle rate for R300 is 100% of clock rate 235MTriangles/sec

triangle rate for NV30 is 350MTraingles at 500Mhz or .7 Traingles per clock

Fill rate: same on a clock per clock basis

Ram bandwidth: R300 has more of it available on a clock per clock basis... what would the bandwidth be if ATi cranked up the clock to 500Mhz DDR? insane that is for sure 31.6GB/sec again TWICE what Nvidia could offer at the same clock.

FSAA quality: Ati wins hands down

FSAA speed: Again ATi wins on a clock per clock basis

R300 16Billion samples/sec vs NV30 15.6Billion samples/sec

Heat: ATi wins R300 runs pretty damned hot but gets by with standard cooling while NV30 uses some form of a mini dustbuster to cool it off.

Noise: ATi will not cause you to go deaf unlike the NV30

Originally posted by Uttar
Hey,
snip

Uttar

[Uttar]

http://www.hardware.fr/medias/photos_news/00/05/IMG0005783.gif

That french site did a comparaison of the GFFX & Radeon 9500 Pro, clock for clock ( both are clocked at 275/270 )
It's Radeon 9500 Pro because it made no sense to compare it with a 256-bit bus, because then you'd have to put the Radeon 9700 Pro at a memory clock of 135, which might hurt latency and give worse performance than normal.

The FSAA/AA benchmarks aren't very fair because ATI uses performance and nVidia uses balanced. As I said before, the most fair thing to do is to use ATI 4x Performance and nVidia 8x Aggressive. It's not highest quality, but it's a VERY fair comparaison. And yes, the aggressive algorithm is too aggressive, I already talked about it in the original post.

However, the no FSAA/AF benchies clearly show the NV30 is slightly ( 5% ) superior clock for clock when compared to the R300.


Uttar

[-=DVS=-]

YeuEmMaiMai

You made a typo Radeon 9700Pro core 325 and it does 325 mill/s triangels not 235 :D

[-=DVS=-]

Originally posted by Uttar
http://www.hardware.fr/medias/photos_news/00/05/IMG0005783.gif

The FSAA/AA benchmarks aren't very fair because ATI uses performance and nVidia uses balanced. As I said before, the most fair thing to do is to use ATI 4x Performance and nVidia 8x Aggressive. It's not highest quality, but it's a VERY fair comparaison. And yes, the aggressive algorithm is too aggressive, I already talked about it in the original post.

However, the no FSAA/AF benchies clearly show the NV30 is slightly ( 5% ) superior clock for clock when compared to the R300.


Uttar

Ati 4x Performance is equal to Nvidia 8x balanced , agressive reduced quality even more :o
Ati's 4x quality is no match at all Nvidia can't do the same

[YeuEmMaiMai]

uhh why would we have to lower the clock of the radeon ram down to 135? GFFX uses DDR-II ram while Radeon uses DDR-1 so the clocks remain the same period


so the GFFX has a slightly better fill rate with no eyecandy but once you enable the eyecandy there is no comparison Ati wins hands down

[Hellbinder]

However, the no FSAA/AF benchies clearly show the NV30 is slightly ( 5% ) superior clock for clock when compared to the R300


I hope I dont get slapped for this.. but this is because the Nv30's Frame-buffer compression is *on* and being used 100% of the time, while Atis is only used during FSAA.

[Bigus Dickus]

Originally posted by Uttar
In case you want the very best quality, you compare nVidia's 8x Balanced with ATI's 8x Performance.
Huh? Wouldn't the very best quality be nVidia's "application" setting (shown at B3D to be different and slightly superior to "balanced") and ATi's "quality" setting?
The other solution is to compare nVidia's 8x Aggressive with ATI's 4x Performance: this gives significantly lower IQ, but the IQ is the same on both cards...
Huh? You must have been looking at different screenshots than I was. Ati's "performance" setting is clearly superior to nVidia's "aggressive" setting in IQ. Perhaps that was just a typo?

[StealthHawk]

Originally posted by Hellbinder
I have a totally different Theory for you on this.

3dMark Fillrate test multitexture

From extremetech.com

GFFX ULTRA 3557.8
9700 Pro 2564.1

From Anand

GFFX ULTRA 3477.9
Radeon 9700 2536.7

Combine those results with the single texture results. The 9700 Pro is clocked at 325 mhz engine. That gives a theoretical maximum of 2600 in the 3D Mark multitexture fillrate test. The actual score is 2564, which is 98.6% of the maximum. Anandtecchs scores are nearly the same. Now look at the GeForce FX. Theoretical maximum at 500 mhz would be 4000, yet it achieves 3558 which is about 89% of the maximum. That seems very odd. Now, take 3558 and round it up to 3600 and now compute the clock speed: 450 mhz. Quite frankly, the multitexture test should be a very reliable way to calculate clock speed.

What conclusion do i draw from this??

By the time that the GFFX gets to the fillrate tests, it is already overheating to the point that the clock speed is getting droped by an automatic 50mhz. But this is just a theory you know. I could be wrong.

If true, it bodes very bad things for those that plan on actually running games with the GFFX.

and i have a different theory for you. the gfFX is bandwidth limited and can't use all of its fillrate.

[legion88]

Originally posted by Uttar
http://www.hardware.fr/medias/photos_news/00/05/IMG0005783.gif

That french site did a comparaison of the GFFX & Radeon 9500 Pro, clock for clock ( both are clocked at 275/270 )
It's Radeon 9500 Pro because it made no sense to compare it with a 256-bit bus, because then you'd have to put the Radeon 9700 Pro at a memory clock of 135, which might hurt latency and give worse performance than normal.

The FSAA/AA benchmarks aren't very fair because ATI uses performance and nVidia uses balanced. As I said before, the most fair thing to do is to use ATI 4x Performance and nVidia 8x Aggressive. It's not highest quality, but it's a VERY fair comparaison. And yes, the aggressive algorithm is too aggressive, I already talked about it in the original post.

However, the no FSAA/AF benchies clearly show the NV30 is slightly ( 5% ) superior clock for clock when compared to the R300.


Uttar

The GF FX uses DDR-II memory. The advantage of DDR2 is that it has the potential to operate at a higher frequency, giving it an advantage over the original DDR-I memory. Since the DDR2 memory of the FX is clocked at the same frequency as the DDR-I memory of the 9500, then the FX no longer has the advantage and is potentially at a disadvantage.

Anyway, it's an interesting idea. The idea is obviously this (for those fanboys too dense to understand even the simplest of things): given the same amount of fillrate and same amount of bandwidth, which card provides the faster performance?

Unfortunately, they messed up that comparison (at least for this graph). If there are other graphs, I haven't seen them.

The two graphics card have two settings for aniso: slow and fast. Just because ATI and NVIDIA decided to call these two settings by different names is irrelevant. The french site decided to run the ATI card in the fast setting and the NVIDIA card in the slow setting.

[The One]

so how would the fx perform haveing a 256 bit bus agaist the r300????

[The One]

i think nvidia should of came out the the gf4 ti 4600 with a 256 bit bus cuz the ti 4600 wasnt to far brhind the r300 with no aa or af once the turn on the eyecandy it got beat bad but it was cuz of the 256 bit bus the r300 has.if nvidia would have gone this route it would of gave them more time to finish and polish the fx

[The One]

i just found this at [h]ard ocp

[H]ardNews 6th Edition Tuesday January 28, 2003
Posted by Kyle 12:28 PM (CST)

GeForce FX Reviews Wrong?
While we are still looking into this, it seems that the in-game screen shots posted on the Net yesterday showing off IQ produced by the GeForceFX 5800 Ultra are "wrong".

There is no doubt that we criticized the GFFX for its AntiAliasing, and now it seems that we may have not had the proper evidence to base our conclusions on. To quote ourselves from this page:


With NoAA you can see the aliasing is quite predominant. 2X AA and Quincunx don’t seem to do much on the GeForceFX visually, but the FPS are affected comparing the shots to the original with no AA enabled.

Of course all of this left us a bit puzzled, and wondering about the AA abilities of the drivers, but the "facts" are the fact correct?

We have been working with NVIDIA on this to get an answer and it seems that now we have the preliminary information to give us a bit more insight on the question.

The GeForceFX's technology applies filters that effect AntiAliasing and Anisotropic filtering before the frame buffer and after the frame has left the frame buffer. In short, this means that all of our screenshots do not accurately represent the true in-game visual quality that the GFFX can and will produce, as the screen shots were pulled from the frame buffer (in the "middle" of the AA process). We have come to conclusions about the GFFX IQ (Image Quality) that may be simply wrong.

While we cannot answer for other reviews of the GeForceFX it is very possible this is an issue with those articles as well, if they were in fact thorough enough to cover IQ.

We are currently working on a way to capture the images properly and will be revisiting the GeForceFX 5800 Preview by covering the IQ portion of our preview with proper screen shot comparison or further information addressing the truth surrounding this situation.

Certainly this is a huge issue it seems that NVIDIA was not even aware of when they issued us the review units. Having 48 hours to preview the card over Superbowl weekend compounded this, and while that is no excuse for improper evaluation on our part, it did certainly impact our ability to do a better evaluation. We are sorry for any incorrect evaluations we have made and are working now to remedy the situation. Any new information will be posted here on our news page.

To answer a popular email, we are currently evaluating in-game play with the screen shots and we will certainly report further on our findings when we have a better handle on the situation.

[mikechai]

The 2xaa/AF image quality issue is currently being discussed in the following thread.
http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=6743
________
Triumph Rocket III (http://www.cyclechaos.com/wiki/Triumph_Rocket_III)

[Uttar]

Originally posted by Bigus Dickus
Huh? Wouldn't the very best quality be nVidia's "application" setting (shown at B3D to be different and slightly superior to "balanced") and ATi's "quality" setting?

Huh? You must have been looking at different screenshots than I was. Ati's "performance" setting is [b]clearly superior to nVidia's "aggressive" setting in IQ. Perhaps that was just a typo?

Well, yeah, it would be "application" settings compared to Quality settings. But I don't really think "application" changes AF at all. Not sure, however.

Yes, ATI's "performance" setting is clearly superior to nVidia's "aggressive" setting in IQ. But, if you compare nVidia's 8x Aggressive and ATI's 4x Performance ( notice the 4x ) , I think the IQ looks about the same.


Uttar