ATI X800 review

[WedgieMan]

http://www.tomshardware.com/graphic/20040504/index.html

/me starts saving up. :bigthumb:

[Despoiler]

A link to a full list of reviews.

http://www.anandtech.com/#22108

All I have to say is that its nice to see the performance race heating up again. Its only good for consumers when the 2 super powers are battling. I still gotta give the crown to ATI because it has better performance with AA and AF. Its weak in the OGL area though. I hope its just some driver optimization that needs to be done.

I am definately shaking the couches out to get one.

[Lollan]

Drool...

[mellotron]

Well the performance in newer games make the new ATi offerings very nice. My only concern is the limitations of games using SM3.0 down the road and how it'll affect performance as well as IQ. Perhaps Ail can shed a little light.

[Jimby]

Bah the NVidia card probably wont be that fast on full PS3.0 anyway since its their first release. By the time PS3.0 is fully used, ATi will have something out that will kick @ss.
I say its a good year and a half before we start seeing some good PS3.0 games.

[babystinky]

Interesting to see these cards on PCI-Express?
I wonder what the performance ratio's will be?

[Mole]

Nvidia who?

:dunno:

[Flywest]

Yeah performance in shaders intensive games and in DX games in general is incredible. Man to check out my card compared to this one on the adnandtech graphs I had to look at the complete bottom. 9700 PRO.

[Apoch003]

AIL shed a little light??? Two weeks ago he was telling everyone to save up money for then next gen. And what did we get? Faster dx9. BAH!!!
Call me when something REALLY interesting shows up. Like a video card that's so fast, it travels time!

[babystinky]

Come on Apoch

You already know all video cards travel time :P

[Mole]

Looks like the 6800U can beat the Radeon X800 series in a whopping 2 tests! Call of Duty and Aquamark.

[Ailuros]

AIL shed a little light??? Two weeks ago he was telling everyone to save up money for then next gen. And what did we get? Faster dx9. BAH!!!
Call me when something REALLY interesting shows up. Like a video card that's so fast, it travels time!

So you would have considered a NV40 if the box would state DX10 on it? You're either kidding me or you don't know what the hell you're talking about.

dx9.0 incorporates more than one graphics generations and you better start getting used to it, since the story won't change with DX-Next.

Yes SM3.0 is one generation ahead of SM2.0 and it could have easily made up for a different API.

In the end any kind of wizzbang feature is worthless for you and only useful for developers; you get twice R360 performance across the board with either R420 or NV40. What the hell else do you want to see?

[Ailuros]

Well the performance in newer games make the new ATi offerings very nice. My only concern is the limitations of games using SM3.0 down the road and how it'll affect performance as well as IQ. Perhaps Ail can shed a little light.

PS3.0 doesn't seem to take much effort to get implemented in current games with PS2.0 or PS2.x routines. The result will be most likely be just performance advantages, to what degree I have no idea yet.

However if a developer today wants to start dealing with extra long shaders and can handle dynamic branching in such a manner that it won't slap back in his face, the developing platform to use for him is definitely a NV40. Longer shaders + branching means hundreds of more instructions and inevitably higher IQ. Presupposition would be to start developing based on those from scratch.

Vertex texturing in VS3.0 and some advanced physics or displacement mapping that's possible through those are probably the most interesting part of SM3.0. Again for developer interest and for those that want to push the envelope in games that'll appear on shelves in more than 2 years from now.

It is important whoever gets to developers first with added features and performance of SM3.0 currently is IMHO less relevant at the moment as long as a developer can get at least 10 fps in 640*480; that's the lowest threshold for "unlimited resources" in DX-Next anyway.

SM4.0 compared to SM3.0 is virtually SM3.0+unlimited resources; not so "unlimited" as you can see above. DX-Next starts introducing other features apart from shaders in the I/O Model, Topology, Tesselation direction. Latter two can be easily treated with a Programmable Primitive Processor (high performance HOS, large hardware cost) or just get some funky implementation through the shaders (low performance checkbox feature HOS, minimal hardware cost); depends on what IHVs intend to do.

If it's another "it's fine as long it's in the featurelist" venture from the big two, then advanced HOS and inevitably true and highly flexible displacement mapping will get once more postponed into the future.

Here's to say that DM was introduced on paper by Matrox as a first hardware based sollution; their EMBM (actually a Bitboys patent AFAIK) had tough luck. It was the only bump mapping method that had dependant texture reads and by the time it got supported in hardware on a wider scale it was rendered redundant by pixel shaders. True hardware displacement mapping might end up having the same luck in the end, since there's no global IHV support for it for the time being.

[Apoch003]

So you would have considered a NV40 if the box would state DX10 on it? You're either kidding me or you don't know what the hell you're talking about.

dx9.0 incorporates more than one graphics generations and you better start getting used to it, since the story won't change with DX-Next.

Yes SM3.0 is one generation ahead of SM2.0 and it could have easily made up for a different API.

In the end any kind of wizzbang feature is worthless for you and only useful for developers; you get twice R360 performance across the board with either R420 or NV40. What the hell else do you want to see?
I thought it was obvious I was kidding. :dunno:

[Ailuros]

I thought it was obvious I was kidding. :dunno:

Couldn't tell; hence the either/or second sentence above.

How bout a smiley once in a while? :)

[Rainbow]

"The power consumption of the X800 XT is about the same as that of its predecessors in 3D applications. Additionally, the cards require only one auxiliary power connector and don't need an especially potent power supply like the GeForce 6800 Ultra does. Even the cooler has shrunk a bit, reducing the card's overall weight and ensuring that it would fit even into a mini-ITX case."

Impressive

[babystinky]

Rainbow

Yeah the ATI solution this time again does seem to be route to go if one upgrades.
The power alone gives a very good reason.

I think the pro version is the card to get per~ ratio/price.

[LaMont Sanford]

"The power consumption of the X800 XT is about the same as that of its predecessors in 3D applications. Additionally, the cards require only one auxiliary power connector and don't need an especially potent power supply like the GeForce 6800 Ultra does. Even the cooler has shrunk a bit, reducing the card's overall weight and ensuring that it would fit even into a mini-ITX case."

Impressive

Meh. Do I have a wimpy psu or case? No. So do I care if ATIs part runs in little junk boxes? No.

As usual, there's nothing wrong and much right about ATIs new cards, but there are definitely other things to consider this time:
1. Will there be PS3 games this year, and what will the difference be?
2. Will Doom3 come out this year, and if so, will it be significantly better on nVidia due to their relationship?
3. Will HL2 see the light of day and be significantly better on ATI?
4. What's up with the current DX9 PS2 benches? Are they running mixed mode on nVidia (maybe bad) or straight FP32 (probably good)?

[incandescent]

1. Will full use of SM3.0 be sufficiently fast on the GF6800? If SM3.0 is used in games, will there be SM2.0 fallbacks, without reducing quality?
2 + 3. ATI & NVIDIA are probably waiting eagerly for these games to hit. Doom3 is going to be a lot of positive press for the 6800, and HL2 is going to be a lot of positive press for ATI
4. If they are mixed mode, as image quality suggests, jumping to full precision could potentially hurt performance for the GF6800, widening the gap even further.

[OtakingGX]

Isn't Far Cry nVidia's heralded SM 3.0 title? I remember reading some news a while back about some ATi guy saying it looked exactly the same as 2.0 code.

Anyway, the X800 Pro slaps the 6800 Ultra around a bit in that benchmark. Then the X800 XT Platinum Edition just bends it over and makes it cry.

[Flywest]

Well, to me , it's clear ATI has won again that round. I don't care about old OGL games based on the Q3 engine, they are fast as hell on the ATI...not as fast as NVIDIA but still those games will eventually dissapear. What's more important is DX9 and PS games, which are alot faster on the ATI...we'll see benchs in Doom 3 how the cards compare themselves.

[LaMont Sanford]

Well, to me , it's clear ATI has won again that round. I don't care about old OGL games based on the Q3 engine, they are fast as hell on the ATI...not as fast as NVIDIA but still those games will eventually dissapear. What's more important is DX9 and PS games, which are alot faster on the ATI...we'll see benchs in Doom 3 how the cards compare themselves.

No kidding? The guy from Canada says it's clear ATI (a Canadian firm) has won even though all benchmarks are pretty close and the nVidia card has more features?
Go figure.

[Flywest]

No kidding? The guy from Canada says it's clear ATI (a Canadian firm) has won even though all benchmarks are pretty close and the nVidia card has more features?
Go figure.


Look like I've hit a sensitive spot .

:ggowned:

[LaMont Sanford]

Look like I've hit a sensitive spot .

:ggowned:


:rolleyes:
Yeah, I'm sensitive to listening to fanATIcs expound about the "clear victories" of their favorite video cards when the differences between them are trifling.

But do tell Flywest, why the X800 series is the best thing to hit the planet since Rush, Molson, and the MapleLeafs....

:rolleyes: :rolleyes: :rolleyes:

[Simian]

Why ATI can say it quite well themselves:

It was very important to us that NVIDIA did not know exactly where to aim. As a result they seem to have over-engineered in some aspects creating a power-hungry monster which is going to be very expensive for them to manufacture. We have a beautifully balanced piece of hardware that beats them on pure performance, cost, scalability, future mobile relevance, etc. That's all because they didn't know what to aim at.


:cool:

[Simian]

No kidding? The guy from Canada says it's clear ATI (a Canadian firm) has won even though all benchmarks are pretty close and the nVidia card has more features?
Go figure.

:lol:

I'll remember to tell the next guy that says an NVIDIA card is superior that he must be a fanboy because he lives in the Unites States. :rolleyes:

[LaMont Sanford]

Why ATI can say it quite well themselves:
:cool:

I suppose that does beat saying: We added some pipelines to our year and a half old design, and hope no PS3 games or Doom 3 come out soon?


Was it Sweeney who said,"Only a marketing person would call 24 bit "full precision""?

[Simian]

Wait, wasn't Farcry PS3? ;)

[OtakingGX]

Wait, wasn't Farcry PS3? ;)That's what I was asking.

From theinquirer.net: (http://www.theinquirer.net/?article=15662)
But it's with Direct3D games like Far Cry - ironically much promoted by Nvidia under its The Way it's Meant To Be Played programme, and being rolled out as one of the first Pixel Shader 3.0 (PS 3.0) games where PS 2.0 Radeon X800XT Platinum shines the most.

I recall certain people saying that while support for new features is nice and all, it's fairly useless seeing as by the time we have titles that really take advantage of them, the cards will be too slow to play those games.

Right now the X800's slight lead in performance is more tangible than the 6800's support for newer shaders. Especially when you consider that the Far Cry example just takes the wind out of that sail.

[RAnta]

Yes you have to take into account all current Dx9 PS2 cards needs to be phased out and that will take atleast 1.5 years by my guesstimations I mean the majority still has Dx8 cards.

So I think one could fairly safely assume that by the time PS3 is mandatory Nv40 and R400 will be quite obselete for consumer all IMHO of course. For devs Nv40 might be prefferable.

Warning Speculations:
Anyhow I have high hope of OC ability for X800Pro really since it is the same silicone as X800PE(XT) it should rather easely OC to 520mhz(XT speeds) wich would yeild an 9.5% increase in fillrate(and ps execution?) thereby closing some of the gap between X800Pro and X800PE.

And Ail would I be correct by assuming that most of the time these beasts will be either be Bandwith(vs fillrate) limited or Cpu limited? Most of the benchies seem to indicate severe Cpu limitations seems like you would need an 3ghz athlon64 to feed them cards.

Anyhow I wonder how the Oc will be like since it seems an artic-cooler will be sufficent to cool R400 cards since power consumtions seems to be at the level of R360.

I would wish that Nv40 was slightly better though that way ATI wich has more headroom for pricecuts (160MILL parts and capaple of disabling quads needs less cooling) would be cheaper than it will be now. Hope Ati will try to undercut NV in price.

[Rainbow]

silicone

:rolleyes:

[DUKE HARDKNOCK]

2. Will Doom3 come out this year, and if so, will it be significantly better on nVidia due to their relationship?
Who cares.

3. Will HL2 see the light of day and be significantly better on ATI?
Who cares.

[Flywest]

:rolleyes:
Yeah, I'm sensitive to listening to fanATIcs expound about the "clear victories" of their favorite video cards when the differences between them are trifling.

But do tell Flywest, why the X800 series is the best thing to hit the planet since Rush, Molson, and the MapleLeafs....

:rolleyes: :rolleyes: :rolleyes:


I don't listen to Rush, I drink Molson beer and I dig their ads , but you're wrong again, I'm a Montreal Canadiens Fan.

Well to me , a silent and not power hungry piece of hardware that does NOT take up to slots in my computer AND manages to beat the competition in 98% of the cases is a "clear winner"...but again, this is IMO...or maybe alot of people share it.

I suppose that does beat saying: We added some pipelines to our year and a half old design

We can talk about GeForce 1, GeForce 2, GeForce 3 and GeForce 4 if you want. :D

[Apoch003]

Who cares.


Who cares.
Me! I have a free copy coming to me because Brian paid x amount of dollars for my video card!

I want my half life 2!!!

[Ailuros]

:rolleyes:
Yeah, I'm sensitive to listening to fanATIcs expound about the "clear victories" of their favorite video cards when the differences between them are trifling.

But do tell Flywest, why the X800 series is the best thing to hit the planet since Rush, Molson, and the MapleLeafs....

:rolleyes: :rolleyes: :rolleyes:


Hmmm the father of all objectivity is back :bigthumb:

(obvious joke disclaimer for the sensitive kind)

[Ailuros]

1. Will there be PS3 games this year, and what will the difference be?
2. Will Doom3 come out this year, and if so, will it be significantly better on nVidia due to their relationship?
3. Will HL2 see the light of day and be significantly better on ATI?
4. What's up with the current DX9 PS2 benches? Are they running mixed mode on nVidia (maybe bad) or straight FP32 (probably good)?

1. If, then mostly performance advantages. There's not enough time for real implementations which means that wherever we see it appearing games are mostly layed out for PS2.x and can get a modest update via extra extensions for more performance with PS3.0. So the real answer is no, since there aren't going to be real SM3.0 games, under the same logic where games like TR:AoD are actually games with a "DX7.0 skeleton" and some PS2.0 effects added to the mix.

2. Doom3 will most likely have a performance advantage and the performance advantage is already visible in stencil fill-rate bound synthetics. Or more simple twice the Z/stencil throughput does make a difference even on paper; and no not thanks to any relationship. If Carmack has actually trashed the NV30 path for Doom3 in favour of the ARB2 path (which according to him hasn't any performance penalty anymore on NV40), then NV3x users are virtually screwed and that irrelevant to any "relationship" too.

3. No. This doesn't mean that it isn't arguable; as arguable as the mixed mode possibly on NV3x based accelerators.

4. Find answer here as an example:

http://techreport.com/reviews/2004q2/radeon-x800/shadermark.gif

http://techreport.com/reviews/2004q2/radeon-x800/index.x?pg=14

NV40 not being able to run all tests though is bad news; hopefully just a driver and not a hardware issue.

Another interesting tidbit:

http://techreport.com/reviews/2004q2/radeon-x800/x800-rthdr2.jpg

http://techreport.com/reviews/2004q2/geforce-6800ultra/6800-rthdr2.jpg

There could lie another point for question (3) since HL2 does make use of HDR lighting. There it looks like rather performance and lower quality for NV40.

For the record ATI will be forced in the longrun to jump to FP32 too sooner or later; the question is whether they'll abandon FP24 entirely or introduce FP16 too since there's not going to be enough bandwidth available for full precision runs under all circumstances for quite some time to come.

[Ailuros]

Yes you have to take into account all current Dx9 PS2 cards needs to be phased out and that will take atleast 1.5 years by my guesstimations I mean the majority still has Dx8 cards.

So I think one could fairly safely assume that by the time PS3 is mandatory Nv40 and R400 will be quite obselete for consumer all IMHO of course. For devs Nv40 might be prefferable.

High end Features have always been to their majority a developer affair and of developer interest; unless someone can show WTF pixel shaders were useful for with the R200/NV20 introductions.

And Ail would I be correct by assuming that most of the time these beasts will be either be Bandwith(vs fillrate) limited or Cpu limited? Most of the benchies seem to indicate severe Cpu limitations seems like you would need an 3ghz athlon64 to feed them cards.

X800XT is far more bandwidth limited than NV40 and the fill-rate scores show it's accelerators fill-rate efficiency (read NV40 and R420 reviews at Xbit-labs for that).

Damn simple:

NV40: 400/550MHz

6.8 GPixels/sec with 35.2 GB/sec bandwidth

R420xt: 520/560MHz

8.32 GPixels/sec with 35.8 GB/sec bandwidth

Yes and as always newer accelerators need also higher speced CPUs to show their real strength. You would had been equally CPU bound in a system containing a 1GHz CPU with a 9700PRO.

Anyhow I wonder how the Oc will be like since it seems an artic-cooler will be sufficent to cool R400 cards since power consumtions seems to be at the level of R360.

Try some liquid cooling for the ram first. The core wouldn't need any significant overclock.

I would wish that Nv40 was slightly better though that way ATI wich has more headroom for pricecuts (160MILL parts and capaple of disabling quads needs less cooling) would be cheaper than it will be now. Hope Ati will try to undercut NV in price.

I still think the 500$ ultra high end models are a waste of money unless they drop in price further down the road. I'd say that most users will be far better off with either a X800PRO or 6800GT, both for price/performance ratios and power consumption.

[LaMont Sanford]

Hmmm the father of all objectivity is back :bigthumb:

(obvious joke disclaimer for the sensitive kind)

Don't forget sensitivity and compassion, Ailuros.
;)

As far as this debate goes, nV40 is a new architecture, AFAIK. I think we'll be seeing a lot of driver fixes and updates over the next few months, similar to when the nV30 came out.
The performance was pretty bad for the specs at first, a month later IQ and speed increased dramatically. I really doubt by the end of July there will be any compelling reason to choose one over the other.

[LaMont Sanford]

2. Doom3 will most likely have a performance advantage and the performance advantage is already visible in stencil fill-rate bound synthetics. Or more simple twice the Z/stencil throughput does make a difference even on paper; and no not thanks to any relationship. If Carmack has actually trashed the NV30 path for Doom3 in favour of the ARB2 path (which according to him hasn't any performance penalty anymore on NV40), then NV3x users are virtually screwed and that irrelevant to any "relationship" too.



Actually Ailuros, Carmack said he scrapped the nV30 path because nV3x cards run the ARB2 path well enough now to do so.

If you don't think Carmack prefers nVidia, you should a. read his plan more, wherein he details how he was developing it on a nv30 due to liking nVidia drivers better, and running into the instruction set limit of the R300 b. notice how he gave reviewers a copy to test the nV30s on? c. remember how ATI leaked the E3 demo
Carmack has always developed on nVidia, AFAIK. I'd call that a preference.

[Flywest]

On a side note, here is an article ( in french sorry ) about the ATI answer to the 6800 Ultra Extreme, which is according to this site again, not a real product, but only an overclocked 6800 ultra.

http://www.3dchips-fr.com/News/actualites/200405077.html

X800XT Platinium Edition overclocked to 715/1278
In short :
3D Mark2003 : 15738 points
UT2k4 1600x1200 AA 4x AF 8x : 147 FPS


ATI did that presentation at the Barcelona's System Builders Summit .

[Rainbow]

Meh. Do I have a wimpy psu or case? No. So do I care if ATIs part runs in little junk boxes? No.

What I was on about - more power consumption = greater case temp = bad.

If I can get a high performance video card that runs at lower temperatures then great, it gets my vote.

[Ailuros]

Don't forget sensitivity and compassion, Ailuros.
;)

As far as this debate goes, nV40 is a new architecture, AFAIK. I think we'll be seeing a lot of driver fixes and updates over the next few months, similar to when the nV30 came out.
The performance was pretty bad for the specs at first, a month later IQ and speed increased dramatically. I really doubt by the end of July there will be any compelling reason to choose one over the other.

There are a lot of issues that weren't "fixed" for the NV3x line for whatever reason. MRTs are one example (actually NV3x supported the "alternative" multiple element textures which were never exposed in drivers) and HDR lighting being another one.

In the infamous HDR from where the pictures are from in the former post, the NV3x's were getting half the framerate of their straigth R3xx competitors and I don't see anything having changed.

Plus an accelerator should be able to run a couple of shader tests; those tests that are missing are pointing in the float textures direction; as I said either hardware or driver issue.

As for the performance dramatically increasing on NV3x; most certainly yes at a sacrifice of quality of course.

[Ailuros]

Actually Ailuros, Carmack said he scrapped the nV30 path because nV3x cards run the ARB2 path well enough now to do so.

If you don't think Carmack prefers nVidia, you should a. read his plan more, wherein he details how he was developing it on a nv30 due to liking nVidia drivers better, and running into the instruction set limit of the R300 b. notice how he gave reviewers a copy to test the nV30s on? c. remember how ATI leaked the E3 demo
Carmack has always developed on nVidia, AFAIK. I'd call that a preference.

The exact comment I've read mentions NV40. You're most welcome to find a link where he claims the NV3x's to have similar performance with partial and full precision.

Development of the Doom3 engine started with the original GeForce from what I recall and the game has been that long under development that I don't even know how many development platforms he might have changed in the meanwhile. Minimum requirement in terms of accelerators will be a NV20 last time I checked.

The reason why he always developed on NV's hardware is that they truly always had the most stable and reliable OGL driver.

However here's your question again:

Will Doom3 come out this year, and if so, will it be significantly better on nVidia due to their relationship?

Now considering my answer, do you really think that NV40 being significantly faster in D3 will be due to any funky conspiracy theory or the simple fact that ATI was just simply too idiotic to not optimize more for stencil performance.

Doom3's major and most resource consuming feature are - as is very well known - it's dynamic shadows. In pure theoretical numbers we have:

NV40@400MHz= 12.8 GPixels/sec stencil fill-rate
X800XT@520MHz= 8.3 GPixels/sec stencil fill-rate

(results borrowed from B3D's NV40 and R420 reviews)

3dmark2k3 Game Test 2:

--------------------------------------1024-------1280-------1600

X800XT--------------------------------90.0-------65.6--------49.8
X800PRO-------------------------------68.2-------48.0--------35.7
6800Ultra-----------------------------94.5-------70.3--------55.0

FableMark:

X800XT-------------------------------158.8-------100.5-------70.3
X800PRO------------------------------126.1--------79.4-------56.2
6800Ultra----------------------------175.0-------110.5-------76.9


More specifically on R420's ROPs:

http://www.beyond3d.com/reviews/ati/r420_x800/index.php?p=9

The basic pixel output engines remain unchanged from R300. In normal operation each pixel pipeline can output 1 colour value and 1 Z/Stencil value, but there is no optimised Z/Stencil rendering performance.

NV40's ROPs:

http://www.beyond3d.com/previews/nvidia/nv40/index.php?p=11

Carmack doesn't need to do NV3x/4x any favours there; ATI is giving the advantage itself for free to the competion.

[LaMont Sanford]

The exact comment I've read mentions NV40. You're most welcome to find a link where he claims the NV3x's to have similar performance with partial and full precision.

Development of the Doom3 engine started with the original GeForce from what I recall and the game has been that long under development that I don't even know how many development platforms he might have changed in the meanwhile. Minimum requirement in terms of accelerators will be a NV20 last time I checked.

The reason why he always developed on NV's hardware is that they truly always had the most stable and reliable OGL driver.

However here's your question again:



Now considering my answer, do you really think that NV40 being significantly faster in D3 will be due to any funky conspiracy theory or the simple fact that ATI was just simply too idiotic to not optimize more for stencil performance.

Doom3's major and most resource consuming feature are - as is very well known - it's dynamic shadows. In pure theoretical numbers we have:

NV40@400MHz= 12.8 GPixels/sec stencil fill-rate
X800XT@520MHz= 8.3 GPixels/sec stencil fill-rate

(results borrowed from B3D's NV40 and R420 reviews)

3dmark2k3 Game Test 2:

--------------------------------------1024-------1280-------1600

X800XT--------------------------------90.0-------65.6--------49.8
X800PRO-------------------------------68.2-------48.0--------35.7
6800Ultra-----------------------------94.5-------70.3--------55.0

FableMark:

X800XT-------------------------------158.8-------100.5-------70.3
X800PRO------------------------------126.1--------79.4-------56.2
6800Ultra----------------------------175.0-------110.5-------76.9


More specifically on R420's ROPs:

http://www.beyond3d.com/reviews/ati/r420_x800/index.php?p=9



NV40's ROPs:

http://www.beyond3d.com/previews/nvidia/nv40/index.php?p=11

Carmack doesn't need to do NV3x/4x any favours there; ATI is giving the advantage itself for free to the competion.
Ailuros:
You know me, Ailuros, at your service:

As for the performance dramatically increasing on NV3x; most certainly yes at a sacrifice of quality of course.
Disagree:
http://www.anandtech.com/video/showdoc.html?i=1821&p=14
Finally, NVIDIA's claim that their "performance" mode offers equal to or greater quality than ATI's is actually true. The benefit here is that NVIDIA actually does some (albeit a small amount) of trilinear filtering in their performance aniso mode, which smooths the transitions between the different mip levels. The tables have turned and now it's ATI's turn to play catch-up and make their performance mode look better.
This was the nV35 review, but you can find in other places how the 5900 drivers raised both IQ and performance for the whole nV3X series in other places as well.

You're most welcome to find a link where he claims the NV3x's to have similar performance with partial and full precision.
This is the Carmack quote I was thinking of, don't remember where I originally saw it.
http://www.beyond3d.com/forum/viewtopic.php?start=80&t=11811
[Doom 3 on the NV40 will use] ARB2. The NV40 is incredibly fast, far and away the fastest thing
you can
get right now.

I have removed the NV30 path, because Nvidia now has the ARB2 path
running
well on the earlier hardware.

John Carmack

I agree with you on the stencil fill rate, but don't think the fact the game is coded in OGl and nVidia traditionally has better OGl drivers hurts either.

I guess it's a matter of preference whether you think his coding the original nV30 path which he said offered similar IQ (I believe on B3d he said the differences would have to be pointed out?) at higher performance constitutes him favoring nVidia or not.
IMO- if a developer goes out of his way to make a game run better on a certain brand, that is favoring that brand.

[Ailuros]

This was the nV35 review, but you can find in other places how the 5900 drivers raised both IQ and performance for the whole nV3X series in other places as well.

I wasn't actually refering to AF (which is another moot point these days since NV sadly adopted angle-dependancy).

http://techreport.com/reviews/2004q2/geforce-6800ultra/5950-rthdr1.jpg

Pretty hm?

http://techreport.com/reviews/2004q2/geforce-6800ultra/index.x?pg=26

Don't bother even NVIDIA internally recognizes that the NV3x was a bad design decision and it's arithmetic efficiency being far from ideal.

The shader compiler definitely helped with mixed modes or partial precisions, but in shader intensive applications it never reached to the day the arithmetic efficiency of the R3xx line.

Anandtech hm?

http://www.anandtech.com/video/showdoc.html?i=2044&p=11

Another interesting read:

http://www.anandtech.com/video/showdoc.html?i=2044&p=4

This is the Carmack quote I was thinking of, don't remember where I originally saw it.
http://www.beyond3d.com/forum/viewt...tart=80&t=11811

The ARB2 path was dog slow in the past on NV3x's because it was forcing it to full FP32 precision; if however NVIDIA managed to get it running at acceptable speed, it means most likely that it's another mixed precision mode and not full floating point values across the application. The parts where really >FP16 are rather small and frankly it's not really worth to use full FP32 across the board. Point acknowledged though, since I've missed that particular comment/quote.

I agree with you on the stencil fill rate, but don't think the fact the game is coded in OGl and nVidia traditionally has better OGl drivers hurts either.

I guess it's a matter of preference whether you think his coding the original nV30 path which he said offered similar IQ (I believe on B3d he said the differences would have to be pointed out?) at higher performance constitutes him favoring nVidia or not.
IMO- if a developer goes out of his way to make a game run better on a certain brand, that is favoring that brand.

Yes the IQ differences according to JC are subtle between FP16 and FP32, because the application rarely requires more than 64bits (see former paragraph).

I do not however believe that he or any other developer makes a game or application on purpose run faster on his "favourite" hardware and screws up the rest.

There are cases where developers use weird rendering paths that might hickup on some hardware, but Carmack and his crew are far from being sluggish developers, rather on the very top of the list when it comes to skill and experience.

Again look back when Doom3 development started (how long has it been now 4 years already or less? in any case a mighty long time), meaning that it has been known for all that time that it heavily relies on dynamic shadows and thus stencil fill-rate/througput being of major importance. ATI didn't have early-Z (which is a requirement for Doom3 rendering since it's an early-Z optimized application) in R200 and both R3xx and R4xx come along with half the stencil fill-rates as NV3x/4x (albeit having early-Z since R3xx).

Would there be a way to switch the stencil ops and thus dynamic shadows off it would run like a champion on about anything out there; here is the major culprit and any possible performance differences are only ATI's fault and not Carmack's or his "preferences". Would ATI have 2Z/stencil units per SIMD and thus 32*0 like NV40, they would in fact have a stencil throughput advantage, since 520*32= 16.6 GPixels/s vs 12.8 with NV40.

It's only a matter of design decisions; ATI didn't only have time for major changes in R420, but obviously set other priorities in transistor counts since R3xx and stencil performance wasn't amongst those immediate priorities.

[incandescent]

i think ATi are banking on HL2 having a better reception than D3. That said, I do not expect the r420's performance to be lackluster in D3 -- it maybe slower than the GF6800, but it's one game. So far, it looks like the r420 is top dog in the majority of benchmarks that count, e.g. farcry, ut2k4, halo, et al.



By benchmarks that count I imply: games with modern feature sets/engine (not Q3 based games, as both the GF68000 & R420 get ungodly fps in those games) and w/ FSAA + AF.

[Sir_Gallahad]

Personally don't think one would be dissappointed getting either card ;) ....until the ramp ups come out in 6 months lol

[Thetargos]

For what I see, it looks like an API preference in the case of both cards: nVidia cards better suited for OpenGL rendering while ATi cards scoring higher in Direct3D benchmarks. For many this actually doesn't matter, since nowadays D3D is the defacto standard for Windows games... But being OpenGL only supported in my OS, that means my upgrading options shrink. I'll see how good/bad do Linux drivers for the next ATi cards (compared to exiting ones) are and maybe wait for middle range nVidia offerings (paying 500+ bucks for a card, I simply can't afford that). I already know what will ATi offer for middle to low range market (being the X800s [Pro and XT] the only offerings now), though it remains to be seen when will they release them (heck we're mearly 24+ hours from the official X800s release :P). So maybe in time these discussions will settle down ( :rolleyes: ), or discussions will only stir them up again ;)

[LaMont Sanford]

Thetargos:
For many this actually doesn't matter, since nowadays D3D is the defacto standard for Windows games
Depends on the games. I have as many or more Quake engine games as I do D3d games. I imagine it will be the same with Doom3 engine games- I play fps, I bet D3 engine will be heavily licensed for fps games.
Even though D3d is the "standard" now, PS2 games that show the advantage of ATI hardware are not, and the gap of their advantage has shrunk by a lot with the 6800s.

[vaaran]

You're most welcome to find a link where he claims the NV3x's to have similar performance with partial and full precision.

ail, heres your link ;)

I'm hoping you can clear up some apparent confusion about DOOM3's rendering paths.

1) There is word that you have removed the NV30-specific rendering path
2) The reason for the above is apparently because NVIDIA's drivers have improved to the point where NV3x hardware are
running the standard ARB2 path at about equal speed with the NV30-specific path

Could you say if the above is true?

Correct.

Also, based on information you provided to the public (via your .plan files as well as your interviews with us),
has there been any significant changes made to the ARB2 path where quality is sacrificed for the sake of performance?

I did decide rather late in the development to go ahead and implement a nice, flexible vertex / fragment programs / Cg
interface with our material system. This is strictly for the ARB2 path, but you can conditionally enable stages with
fallbacks for older hardware. I made a couple demo examples, and the artists have gone and put them all over the place...

What would be the best way to benchmark the game on various hardware? This is actually quite a problem for a site like ours.
Given that there are different rendering paths as well as possibly drivers doing difficult-to-verify call traces
(perhaps some shader replacements and all those sorts of things), how would we be able to present comparable performance
data analysis amongst different hardware? Obviously there are two ways to look at this : one would be from the angle of
gamers who are looking to upgrade their pre-DX9 video cards, another would be for those who are already on a DX9-class
video card that may be tempted to change to one that runs the game better than the one they have.

Dumping the NV30 path makes this much easier. All the cards anyone is really going to care about benchmarking will use the ARB2 path.

http://www.beyond3d.com/forum/viewtopic.php?t=12006

[LaMont Sanford]

Thanks vaaran, I knew I had seen that somewhere.

[Ailuros]

ail, heres your link ;)



http://www.beyond3d.com/forum/viewtopic.php?t=12006

Okeyyyyyyyyy....

I did decide rather late in the development to go ahead and implement a nice, flexible vertex / fragment programs / Cg
interface with our material system.

Not a very clear answer, with the only other difference that it means that the ARB2 path today is not obviously using the exact same settings as in the past. In the past it was forcing FP32 across the board.

That's actually IMHO the NV30-path + forcing to run full precision in rare selected spots.

[Ailuros]

i think ATi are banking on HL2 having a better reception than D3. That said, I do not expect the r420's performance to be lackluster in D3 -- it maybe slower than the GF6800, but it's one game. So far, it looks like the r420 is top dog in the majority of benchmarks that count, e.g. farcry, ut2k4, halo, et al.



By benchmarks that count I imply: games with modern feature sets/engine (not Q3 based games, as both the GF68000 & R420 get ungodly fps in those games) and w/ FSAA + AF.

Let's call them all benchmark key applications to make it easier (albeit I hate the fact that Halo makes it into that list ROFL).

Doom3 and the game titles using it's engine in the future will belong to the key applications inevitably and judging from the Q3a engine and it's licenses it wasn't exactly a small list of popular games.

It would have blown up ATI's transistor budget exactly how much to just double the amount of Z/stencils per clock?

I am not willing to justify any excuses as long as it was possible; I won't excuse if NV's hardware isn't in fact able to expose float textures as it should, as much as I won't excuse lacklustering stencil performance with ATI, even if the ballpark isn't as big (currently the difference seems to be at about 10% tops; that though with 400 vs 520MHz clockspeeds; which simply could mean that NV has a much higher headroom for future increases).

Think of it the other way around: would there had been 32*0 per clock, the X800XT would have had >30% advantage in that department.

All things pulled through an average, it looks like tables have turned in clock for clock performance. In the past it looked like that:

NV30@500MHz (128bit) vs R300@325MHz
NV35@450MHz vs R350@380MHz (or was that 360? too bored to look it up)
NV38@475MHz vs R360@412MHz
.................

6800U@400MHz vs X800XT@520MHz

What's more important here is that the NV40 doesn't seem to have any trouble with it's fill-rate/bandwidth efficiencies.

Efficiency has always been a highly important point for me; here the NV40 isn't only fast but more feature-complete too beyond doubt. As for the ultra high end rare models for the extravagant 1-2% tops out there, it's more a marketing hurdle for the idiotic performance crown. The real high end battle will be between 6800GT and X800PRO.

Not only does NVIDIA have more efficiency per clock but they start out at a much lower clockspeed too and there they have a lot more chances and headroom for future incarnations. Of course could someone easily tell me that ATI could pump clockspeeds way beyond 600MHz this year, to what I'll probably will not get an answer for is whether they'll find enough high speced GDDR3 to feed it with the bandwidth it actually needs this year.

Or is there going to be 800MHz GDDR3 this year in sufficient quantities after all?

[Plaster]

I have no idea what all those settings mean, but I matched them as close as possible. :bigthumb:

[Ailuros]

Thetargos:

Depends on the games. I have as many or more Quake engine games as I do D3d games. I imagine it will be the same with Doom3 engine games- I play fps, I bet D3 engine will be heavily licensed for fps games.
Even though D3d is the "standard" now, PS2 games that show the advantage of ATI hardware are not, and the gap of their advantage has shrunk by a lot with the 6800s.

JC and iD seem to be truly the last frontier when it comes to OGL. The majority of developers use D3D these days.

I think that even CroTeam have jumped on the D3D bandwagon too; which of course is going to be another key application and will hopefully be released this year *drools*. For that both HL2 and Doom3 can kiss my <censored> :naughty:

[LaMont Sanford]

JC and iD seem to be truly the last frontier when it comes to OGL. The majority of developers use D3D these days.

I think that even CroTeam have jumped on the D3D bandwagon too; which of course is going to be another key application and will hopefully be released this year *drools*. For that both HL2 and Doom3 can kiss my <censored> :naughty:

If JC and iD are the "last frontier" of OGl, I'd trade all other games for the games iD has produced, and those based on it's engines. IMO OGL runs smoother and often looks better, or at least as good.

CroTeam?!?!?!
Wow. I only bought SS1. Liked the style of gameplay, but the environments bored me out of plying it much. Painkiller is a game like SS should be, creepy.

[Ailuros]

If JC and iD are the "last frontier" of OGl, I'd trade all other games for the games iD has produced, and those based on it's engines. IMO OGL runs smoother and often looks better, or at least as good.

CroTeam?!?!?!
Wow. I only bought SS1. Liked the style of gameplay, but the environments bored me out of plying it much. Painkiller is a game like SS should be, creepy.

What iD has produced or licensed? :P

Let's see Quake3 Arena was abysmally fill-rate limited for it's time and all other games based on it's engine ended up quite CPU bound with no exeptions I could think of.

SS:SE is times ahead in graphics glory compared to SS1 and hate me for it but the underlying engine was/is at least capable of real multitexturing and not some funky >dual texturing force me to multipass crap of the Q3a engine. Furthermore the launch price for both games was about half of what the rest of the games usually cost.

Developers these days prefer Direct3D because it's easier to code for and they get effects times earlier than with OGL. That's actually the responsibility of the ARB board because it takes too long for them and the IHVs of course to get their heads out of their butts and incorporate features in time. OGL2.0 which is most likely going to incorporate analogue functionalities of the DX9.0 SM3.0 had been postponed again and I guess we're lucky if we see it surfacing in H2 2004.

OGL running smoother than D3D is a myth that belongs to the past; if there was ever such a thing then it was limited to NVIDIA hardware only, because their DX7 D3D performance was rather pathetic; thank God some games like UT had the blessing to get an OpenGL renderer. If an engine has been coded out front for a specific API then it's natural that it'll work with that API better; else try UT2k3 as an example in OGL.

Painkiller was fine yet nothing I hadn't seen before.

Doom3 is basically a game with a DX7-T&L similar skeleton with dynamic shadows and a few spare dx8/9.0 style effects added to the mix.

Serious Sam 2 is a totally revamped engine and as I said hopefully we'll see it launching before this year runs out. I want a game that would make me want to upgrade, from the stuff I've been able to see so far only SS2 qualifies for it; the rest is already playable on current hardware.

[vaaran]

to come back to the stencil performance,
theoretically while using fsaa the numbers would be:
NV40@400MHz - 12.8 gigasamples/sec color/stencil/Z fillrate with FSAA
R420@520MHz - 16.6 gigasamples/sec color/stencil/Z fillrate with FSAA

i just wonder why the performance numbers, in 3dmk03 gt2 for example, indicate otherwise....

from the b3d reviews, gametest 2, 1600x1200, no af

r420
http://www.beyond3d.com/reviews/ati/r420_x800/chart/image031.gif

nv40
http://www.beyond3d.com/previews/nvidia/nv40/chart/image025.gif

[vaaran]

demirurg @3dcenter suggested more bandwidth limitation on the r420 but i still dont get it...

[Ailuros]

demirurg @3dcenter suggested more bandwidth limitation on the r420 but i still dont get it...

Stencil fill-rate is not de-coupled from pixel fill-rate on IMRs. R420 shows due to it's bandwidth limitation a much lower fill-rate efficiency than NV40.

I don't know if that's at fault here, but I guess one way to find out, would be to get one of those liquid cooled-whatever VPUs to run at default clockspeeds with >700MHz ram speeds to see if anything changes.

I personally am counting stencil fill-rates w/o MSAA on purpose. Real time performance isn't suggesting a 30%+ advantage for R420 like those fill-rates do, rather the exact opposite.