Anand's 9800XT and FX5950 review, part 2

[DaveBaumann]

How long after their PR event?

They didn't hide it. I think Craig Peeper from MS may have even mentioned it offhand in his presentation.

So then the R2x0 was cheating? Whenever AF and trilinear were selected, you didn't get it.

You can call it cheating if you want Ben. This was a known architectural limitation that ATI told us the caveats about its implementation ( http://www.beyond3d.com/articles/ects2001/index3.php ), this isn't something that the chip can do but selectively doesn't on some application to uncrease performance. If you'd asked about the dodgy Trilinear filtering in Q3 when R200 was first about, then that goes in the same pot as what has been occuring with NVIDIA's filtering.

You said they wanted to demonstrate the relative performance of ATi v nV in terms of shader performance. Are you now saying that you didn't mean that?

They did for "full precision" DX9.

Then why bother replying? Civility is certainly escaping you here.

Not really, but I'm still illustating the point that despite the fact that you may feel its irrelevent, it is still a valid data point.

HL2 showed a 70% rift which you have been staunchly supporting.

I'm not supporting any number such as 70% - where has this come from? I'm not supporting any such 70% figure that you've pulled out, I'm pointing out what was occuring at that presentation as I was there, I'm also showing why what they showed was valid at the time - that doesn't mean it won't change. We have several other datapoints that also show similar trends. When I said "so pendenant on those variables then the performance difference can be very large or very small." that wasn't in relation to any "70%" figure, but in relation to each architecture

Any developer that did that for a game with the intention of slowing it down I would consider to be extraodinarily stupid, yet it isn't supposed to be poorly coded by your standards.

?

Ask Gearbox about it.

I figured you had, since you put the statement forth...

Performance drops decently on ATi and nV parts enabling the PS 2.0 shaders, just nowhere near the catastrophic gap that you seem to want everyone to think will happen if they try and run a PS 2.0 game on nV hardware.

"I seem to want people to think" - You're going around putting words in peoples mouths again Ben, thats not very becoming now, is it. You;re the one who seems to think I;ve said such things when I have not.

As I said, these things depend entirely on the composition of the shaders, and for what quantity of pixels they run on. I've seen some PS2.0 coded shaders that run faster on the FX.

I will ask you then, what level of performance difference do you expect between FX and R3X hardware when running DX9 games roughly?

That is a foolish thing to ask because its completely unquantifiable, without the data to hand. Some games may make use of lots of DX9 shaders making the difference large (TR, 3DMark, HL2 Full Precision), others may make only small use of it and much of the rest of the scene may be comprised of elements that favour the FX architecture better thus overshadowing the deficits. I'm all likelyhood I'll wager that the true gaming difference will not be too great in most situations while the FX architecture is around.

 

[jbirney]

Dam I lied. I thought I could stay out of this thread. Anyways a few points:

Halo doesn't show anything remotely like HL2. The performance gap using the old drivers is comparitively mild, not to mention it reverses itself with the latest drivers(which are now available for public scrutiny).

From ETs own article on Halo Benchmarking:
"Halo does utilize partial-precision hints, allowing GeForce FX cards to operate using faster 16-bit floating point operations."

Thus your comparing Apples to oranges. If you look you will see that the Halo Benches are more akin to the HL2 mixmode benchies which show almost the same thing. From the benchmarks I saw, HL2 ran an in mixed precesion mode the FX ultra got around 48 compared to ATI 9800pro 62. Which is close to the same difference we saw in Halo PC benchies.

Looks like this will be a trend with all games using FX: From James Loe lead Tec over at Gas Powered Games:

"Precision has a pretty big impact on the current nVidia chips, so it?s important to use half precision data types and the nVidia specific 2.0a HLSL compiler when running on NV3X hardware. Keeping those limitations in mind, it hasn?t been too difficult for us to maintain an acceptable level of performance"

The reason I bring this up? Is becuase in most games you will not have NV using a straigh PS2.0 path. Thus when YOU try to compare them to HL2 benchmarks that used the straight PS2.0 path you will not be looking at the same thing.

Besides why should one game show me what another game does? Why if I get 100 fps in UT2k3 am I lucky to break 50s with splinter cell? Different games, different bottlenecks, different results. Never ever try to compare two dis-simular things and draw up a conclusion based off that comparison.

Your ignoring what I have actually said became tiring some time ago. Yet again, in this thread, I will say that I expect the R3X0 to be in the area of 25%-35% faster then the comparable FX part in PS2 performance on average. HL2 showed a 70% rift which you have been staunchly supporting.

TR:AOD shows a simular thing:
http://www.gamersdepot.com/hardware/video_cards/ati_vs_nvidia/dx9_desktop/tomb_raider.gif


Why only 25% faster? The R350 has 2x the number of pixel engines that are fully dedicated vrs the FX that have to multiplex their time with other functions. I remember a post on B3D forums from the presentations done at Shader Day. Senric (SP????) and ATI employee gave out some of the number of pixels they can push through the R3xx shaders. It was pretty big number but I could not find the info (dam my searching skills). Maybe Dave can find it and post it here vrs what the FX is cable of?

I consider 3DM2K3 to be utterly irrelevant for numerous reasons, mainly it is far too slow on all the DX8 tests versus what you see on screen(particularly GT3, it looks to be sub par to D3 or even Amp2 and yet runs significantly slower). And also their bench wasn't mapped to the FX properly either, although that is most certainly not their fault as the compiler was not ready when they released it which was not the case with Valve.

Then what about ShaderMark and RightMark. They show the same delta in preformance. Any why does it have to be mapped to the FX? It was mapped to the DX9 spec which it was designed to test. Not FM fault if other companies chose to different hardware implementation.

 

[Pete]

Ben, just to be clear, the HL2 benches from ATi's Shader event were based on a build of the engine from E3? That seems a bit far back, but also not very specific. Was the engine code frozen, and most development focused on gameplay/art? Are shaders considered part of the engine, or separate (as Gabe said all their FX-specific work wouldn't translate to other mods/uses of Source)?

BTW, I believe this thread may take the record for most quotes per post.

 

[spam]

Hi Ben,

I would like your clarification on a couple of questions I have.

1. Do you think that the graphics industry and the consumer would benefit by the standardization of API's?

2. What do you think would encourage the graphics hardware manufacturers to move in this direction?

My thoughts are that such a move would help to level the ice on the hockey arena. It would make apple to apple comparisons possible. It would force companies to strive for conformity rather than being proprietary.

3. Do you think your postings in this thread are helpful to achieving that end?

 

[Pete]

What do you mean by "standardization of APIs?" There are already two standards, and DirectX is probably the most-used one by virtue of Windows offering the largest consumer base. A large issue since the FX's intro is whether MS was correct in calling the FX "DX9" at all, given its initial performance. The question is, who's to blame: MS for inclusive definitions (the 5200/5600 are DX9-compliant, but they appear to be too slow to be useful with DX9 effects), or nV, for *really* slow initial drivers and "mis-aimed" initial DX9 hardware (5200/5600/5800)?

It's tough to force all graphics makers to make hardware with the same underlying architecture, as it would sort of make this a race to find the best factories and process improvements, rather than one to find the best achitecture. Would design be by committee? Would we never have seen TBR cards? OTOH, DX9 has proven (at least initially) that it's also tough to shoehorn different but similar architectures with different but similar goals into a single standard, like DX9. I think competing hardware is advantageous, but obviously the 3D industry has a ways to go with software that hides competing hardware's quirks/features/advantages/disadvantages.

Ultimately, if you want "standardization of API's," a console is your best bet.

 

[spam]

I am better at asking questions than answering them but to clarify what I was asking. ....

From what I have read, "please don't make me look it up!" Nvidia wanted MS to conform DX9 to it's own preferences and it is no surprise that MS said no. Nvidia went it's own way in the development of it's hardware. Apparently it made some bad hardware design choices that did not handle shaders according to the MS spec. for the Dx9 API. API's offer more than one hardware design to work effectively. ATI chose 24bit precision which was DX9 compliant, NV chose 16bit and 32bit precision which was not optimal design for DX9 compliance. If Nvidia had chosen to comply to 24bit standard there would not be the difficulty we currently face in apples and oranges comparisons between 16 bit and 24 bit performance.

 

[spam]

It's tough to force all graphics makers to make hardware with the same underlying architecture, as it would sort of make this a race to find the best factories and process improvements, rather than one to find the best achitecture.

There is no need to have the same underlying architecture, race cars designers have their own solutions on how to cross the finish line first.

 

[Genx87]

Actually from what I gathered Microsoft was leaning towards FP16 but ended up going to FP24 sometime after July 12th of 2002.
Nvidia got stuck with thier pants down. Some people have said the standard was made in the Winter of 2001 but I did a little research and many people believed it was going to be FP16 upto at least July of 2002.

 

[spam]

 

[spam]

Adding to my race car design analogy, you also have drivers who use different tactics to cross the line first . Some even cheat.

 

[BenSkywalker]

You can call it cheating if you want Ben. This was a known architectural limitation that ATI told us the caveats about its implementation ( http://www.beyond3d.com/articles/ects2001/index3.php ), this isn't something that the chip can do but selectively doesn't on some application to uncrease performance. If you'd asked about the dodgy Trilinear filtering in Q3 when R200 was first about, then that goes in the same pot as what has been occuring with NVIDIA's filtering.

A better example is the current AF hack they are using in terms of trilinear only on one mip level. Whether or not you select trilinear in game you aren't getting it. I do NOT consider these cheats, I consider them hacks, but I think if you are going to deem one company as a cheater for filtering hacks then the other should be too.

They did for "full precision" DX9.

So not using the FX compiler while trying to display the performance gap was fair?

I'm not supporting any number such as 70% - where has this come from?

With your support of Valve's benches. There are large groups of people on these forums that are using that as what is to be expected in typical DX9 games. Your comments up to your last post have done nothing in the least to change that view.

I figured you had, since you put the statement forth...

They have stated repeatedly that the only way to enjoy the game with all features on is with a decent performing DX9 card with PS 2.0 support. They have also made comments that you don't want to run PS2.0 on the FX5200 because its shader speed is too slow.

You're going around putting words in peoples mouths again Ben, thats not very becoming now, is it. You;re the one who seems to think I;ve said such things when I have not.

I stated 'seems', now you have come out and stated that the 70% rift is not something you expect to be the norm, I no longer would say it seems that way.

As I said, these things depend entirely on the composition of the shaders, and for what quantity of pixels they run on. I've seen some PS2.0 coded shaders that run faster on the FX.

Obviously. This forum has been crawling with people who are convinced that the 5900U is going to be showing sub 30FPS performance in any game that utilizes PS 2.0. I've been sitting back listening to them go on about it for some time. They have been convinced that the only way nV could get close was by cheating an enormous amount and replacing all of their shader code with DX8 level shaders. Now we finally have some drivers that clean up the bugs they had and I've been trying to get across that the DX9 performance of the FX is not going to be nearly as bad as Valve made it seem with their benchmark.

That is a foolish thing to ask because its completely unquantifiable, without the data to hand.

I'm all likelyhood I'll wager that the true gaming difference will not be too great in most situations while the FX architecture is around.

The second quote is what I was looking for. Obviously I'm not going to hold you to a theoretical generalization based on factors that are not given yet, I'm trying to see if you think the enormous performance gap will be close to what many here think it will be in any game that uses PS 2.0.

JB-

Thus your comparing Apples to oranges. If you look you will see that the Halo Benches are more akin to the HL2 mixmode benchies which show almost the same thing. From the benchmarks I saw, HL2 ran an in mixed precesion mode the FX ultra got around 48 compared to ATI 9800pro 62. Which is close to the same difference we saw in Halo PC benchies.

Not quite, the shaders in Halo have no benefit running in full precission. Valve spent some time talking about the reduced IQ in HL2.

Looks like this will be a trend with all games using FX: From James Loe lead Tec over at Gas Powered Games:

"Precision has a pretty big impact on the current nVidia chips, so it?s important to use half precision data types and the nVidia specific 2.0a HLSL compiler when running on NV3X hardware. Keeping those limitations in mind, it hasn?t been too difficult for us to maintain an acceptable level of performance"

Had to add emphasis on that part, it's something I've been saying for some time.

The reason I bring this up? Is becuase in most games you will not have NV using a straigh PS2.0 path. Thus when YOU try to compare them to HL2 benchmarks that used the straight PS2.0 path you will not be looking at the same thing.

With HL2 you end up with different visuals, if you don't then using the Apples to Oranges comment isn't quite valid as running in full precission nV is besting ATi. As long as their is an IQ difference then I think running them in the same path, without PP, is valid. If they are producing the same IQ then it is based on if you want to make ATi look better.

Different games, different bottlenecks, different results.

Dave has been vocal, along with most of the people in this forum, that PS 2.0 are going to be the main limiting factor of DX9 titles in terms of performance. I'm simply going along with that assumption. Obviously D3 shows something very different then HL2, but that is being nigh completely ignored.

Why only 25% faster? The R350 has 2x the number of pixel engines that are fully dedicated vrs the FX that have to multiplex their time with other functions.

They have differing levels of functionality and also the NV3X parts are clocked higher. As Dave mentioned in his last post, certain shaders are faster on the FX.

Then what about ShaderMark and RightMark. They show the same delta in preformance. Any why does it have to be mapped to the FX? It was mapped to the DX9 spec which it was designed to test. Not FM fault if other companies chose to different hardware implementation.

ShaderMark doesn't any more, not sure about RightMark. Why should it be mapped to the FX? Why shouldn't it be? You and I both know that ATi's proposal won for the DX9 spec, it is already mapped to their hardware. If we want to use that same line of thought, then PS 1.4 should have never existed

Pete-

Ben, just to be clear, the HL2 benches from ATi's Shader event were based on a build of the engine from E3?

That's what Dave said. He's talked to the guys at Valve recently while I never have(at least, not as far as I know, maybe over at Evil's but if I did it wasn't about that ).

Spam-

1. Do you think that the graphics industry and the consumer would benefit by the standardization of API's?

No. Game developers could benefit, but not the companies nor really the consumers in the end. If you had to wait for full API specs to be released before you started working on your parts we would be looking at an enormous rift between feature implementation in the API and implementation in hardware. Also, with standardizing the API we either use OpenGL and drop DX which now dominates the gaming market, or we drop support for Linux and Macs, neither of which most people want to see. The one advantage would be that we could see game support for new features sooner in relation to when it becomes available in hardware. If MS/ARB were to try and move quickly getting their spec out years in advance they may under or overestimate where build processes are going to be and hence cause issues with designs(this happened to nV anyway with the NV30, but it could be a lot worse if they ended up being off by a year or more).

2. What do you think would encourage the graphics hardware manufacturers to move in this direction?

If they become the defacto standard. When nV had everything going their way on the API front you didn't see them pushing for implementations to DX to help them out with their architecture, likewise now with ATi.

3. Do you think your postings in this thread are helpful to achieving that end?

Absolutely not, nor is it something I think would be entirely deisrable. My posts in this thread started off pointing out that nV was not nearly the whipping boy a lot of people are trying to claim they are. Relative parity at the high end is something I find highly desireable and is something I think we have right now, despite the objections to that point by others.

 

[DaveBaumann]

A better example is the current AF hack they are using in terms of trilinear only on one mip level. Whether or not you select trilinear in game you aren't getting it.

Wrong. The only time you don?t get it is if its applied through the control panel. If the application selects Trilinear Anisotropic Filtering the Trilinear Filtering is applied to all layers ? since this was discovered I?ve been using application control AF in UT2003 in all the reviews; compare the performance difference for the Radeons before and after the 9800 256MB review. (AFAIK there is also a reg entry to enable it as the ?quality? setting, but that?s by-the-by).

So not using the FX compiler while trying to display the performance gap was fair?

If they are attempting to highlight the path they think many of the developers with lesser resources are going to take, then again this would be what they were trying to illustrate.

With your support of Valve's benches.

I support the data I have in front of me. I?ve seen similar things with TR:AoD, 3DMark, ShaderMark, Marko?s Fillrate tester ? I?ve also seen cases where PS2.0 is in operation and the performance difference is much lower, but thus far they have all been with a much lower utilisation of PS2.0. I?m looking at something else right now and its still similar story. Also, I?ve been looking at ShaderMArk, which allows compilation via PS_2_a and the performance differences are negligible in most cases.

They have stated repeatedly that the only way to enjoy the game with all features on is with a decent performing DX9 card with PS 2.0 support.

I was talking about this statement that you made:

Gearbox has talked decently about how many of the shaders they redid to improve visuals using DX9 shaders. And Halo has plenty of levels that have nigh every pixel running a shader too.

Presumably you have something to back that up?

Now we finally have some drivers that clean up the bugs they had and I've been trying to get across that the DX9 performance of the FX is not going to be nearly as bad as Valve made it seem with their benchmark.

Valve?s benchmark clearly showed that the performance wasn?t greatly different, when using the mixed mode. The performance differences shown are when utilising full precision rendering ? its entirely down to the developer whether they do that or not, and on what quantity of the screen they deem it necessary.

Obviously D3 shows something very different then HL2, but that is being nigh completely ignored.

D3?s bottleneck is not shader driven to any great extent. The Bottleneck is the stencil shadow performance.

ShaderMark doesn't any more, not sure about RightMark. Why should it be mapped to the FX? Why shouldn't it be? You and I both know that ATi's proposal won for the DX9 spec, it is already mapped to their hardware. If we want to use that same line of thought, then PS 1.4 should have never existed.

ShaderMark2.0 does show a fair performance delta, and that?s with the PS_2_a compiler.

That's what Dave said. He's talked to the guys at Valve recently while I never have(at least, not as far as I know, maybe over at Evil's but if I did it wasn't about that ).

Did I?

The levels are from the E3 demo, that says nothing about the build of the engine.

 

[BenSkywalker]

Wrong. The only time you don?t get it is if its applied through the control panel.

I was only talking about the 99% of games that did not have an option to select AF. If you think that is only older games, Halo doesn't have the option.

since this was discovered I?ve been using application control AF in UT2003 in all the reviews

Of course, how many others do you think are?

If they are attempting to highlight the path they think many of the developers with lesser resources are going to take, then again this would be what they were trying to illustrate.

Not compiling the code for the market leader...? Why would a developer with lesser resources ignore the largest market?

I?ve seen similar things with TR:AoD, 3DMark, ShaderMark, Marko?s Fillrate tester ? I?ve also seen cases where PS2.0 is in operation and the performance difference is much lower, but thus far they have all been with a much lower utilisation of PS2.0.

But I have not questioned those(well, I think 3DM2K3 is poorly done, trying to cripple vid cards for a supposed game test..?). I don't doubt that there will be a performance difference, the point I've been staunchly opposing is that the 70% figure is close to what we will see as the norm.

Presumably you have something to back that up?

Play the game. I didn't say that some levels have nearly every pixel covered by a DX9 shader in particular(although I suppose for collapsing passes some may have been promoted) but they certainly had shaders running everywhere in certain levels(you can also disable shaders yourself and play through them to take a look, actually you can force fixed function, PS 1.0, 1.1, 1.3, 1.4 or 2.0).

D3?s bottleneck is not shader driven to any great extent. The Bottleneck is the stencil shadow performance.

And the common theme here is that everyone is going to jump on the HL2 engine and noone would dare use Doom3 because of the fact that it doesn't use PS 2.0. I know you aren't saying it(though you have been supporting the widespread PS 2.0 penetration soon), but that is the illusion that many people are giving to people. HL2 is all that matters, Doom3 is DOA.

ShaderMark2.0 does show a fair performance delta, and that?s with the PS_2_a compiler.

Not 70% on average, not even close to that rift which is the point I've been taking issue with all along.

Did I?

The levels are from the E3 demo, that says nothing about the build of the engine.

HDR had only just been implemented. The benchmark was from the E3 levels/build and HDR had not been implemented in them.

Yes, you did say it was from the E3 build and expanded to state that is why HDR was not used as it had not been implemented at that point.

 

[jbirney]

Not quite, the shaders in Halo have no benefit running in full precision. Valve spent some time talking about the reduced IQ in HL2.

I really don't understand this comment?

Let me restate the point:

HL2 in full precision mode showed us a 70% delta between the two.

You pointed out that Halo did not show same delta.

I pointed out that Halo does not have a full precisions mode and uses partial precision for FX

Thus my point comparing Halo numbers to HL2 full precision is pointless. What's to ague about that? They are Full Precision to Partial precision of course there going to be different. Why would other games show the same thing when they don't do the same work?

Had to add emphasis on that part, it's something I've been saying for some time.

The new shader mark allows you to compile with this hint and showed no real increase.
Edit: doh see Dave posted that already but see numbers below

As long as their is an IQ difference then I think running them in the same path, without PP, is valid. If they are producing the same IQ then it is based on if you want to make ATi look better.

John C has already said their IS an IQ difference in D3 between the two. Its gonna depend on the game I feel.

Dave has been vocal, along with most of the people in this forum, that PS 2.0 are going to be the main limiting factor of DX9 titles in terms of performance. I'm simply going along with that assumption. Obviously D3 shows something very different then HL2, but that is being nigh completely ignored.

Your smart enough to know there is a difference between looking at different games too draw a specific conclusion vrs looking at many games test and drawing a general conclusion. The general conclusion looking at TR, HL2, Halo, Shadermark, 3dmark, Rightmark,ect is that the R9800 is much faster running PS2.0 shaders. I don't think anyone here will disagree with that. And before the Det50s, some benches showed much more than 70% faster which sorta backs up the HL2 numbers. TR numbers showed about 80% if you bothered to look.

I am sure PS2.0 will be come more limiting factor. But you have to remember not every game will use them in the same manor. AQ3 for example only has one PS2.0 shader. HL2 had a ton of PS2.0 shaders.

They have differing levels of functionality and also the NV3X parts are clocked higher. As Dave mentioned in his last post, certain shaders are faster on the FX.

So in fact you don't really know and its just and educated Guess. Again I will try to find the numbers from the ATI presentation.

ShaderMark doesn't any more, not sure about RightMark.

ShaderMark still shows about a 40% different between the two.

http://homepage.ntu.edu.tw/~b91602047/smtemp.gif

Found more numbers with your PS2.0A stuff

Radeon 9800 Pro, stock and Radeon Pro 256 MB (OVERCLOCKED 405/371: ), Catalyst 3.7 drivers (Cheat detection enabled / disabled...no difference)
GeForce 5900 Ultra, Det 52.xx, PS 2.0a, partial precision mode.

See this thread for more numbers: http://www.beyond3d.com/forum/viewtopic.php?t=8193&postdays=0&postorder=asc&start=20


Test Delta (5900U vrs 9800Pro)
-----------------------------------------------------
2 -33.2%
3 -31.3%
4 (N/A)
5 -30.8%
6 -27.6%
7 2.7%
8 (N/A)
9 -47.4%
10 -31.5%
11 -31.6%
12 -36.0%
13 -26.4%
14 -6.9%
15 -32.0%
16 -21.1%
17 -23.1%
18 -46.7%
19 (N/A)
20 (N/A)
21 (N/A)
22 (N/A)
23 (N/A)

Notice the how many test the FX can not even run. Shall we factor those in

Why should it be mapped to the FX? Why shouldn't it be? You and I both know that ATi's proposal won for the DX9 spec, it is already mapped to their hardware. If we want to use that same line of thought, then PS 1.4 should have never existed

Does it make more sense to map your benchmark to the DX9 standard or map it to something non-standard?

 

[Genx87]

Does it make more sense to map your benchmark to the DX9 standard or map it to something non-standard?

Standard or not it would make a lot of sense to map it to hardware that is the majority of the market.

 

[spam]

Ben,

No. Game developers could benefit, but not the companies nor really the consumers in the end. If you had to wait for full API specs to be released before you started working on your parts we would be looking at an enormous rift between feature implementation in the API and implementation in hardware. Also, with standardizing the API we either use OpenGL and drop DX which now dominates the gaming market, or we drop support for Linux and Macs, neither of which most people want to see. The one advantage would be that we could see game support for new features sooner in relation to when it becomes available in hardware. If MS/ARB were to try and move quickly getting their spec out years in advance they may under or overestimate where build processes are going to be and hence cause issues with designs(this happened to nV anyway with the NV30, but it could be a lot worse if they ended up being off by a year or more).

You mentioned, Macs, Linux , DX and OpenGl having different deisign requirements . This is a very complex design problem for companies that want to sell to these respective markets.

-Departures from these API's(OpenGl and Direct3d) would only add to the difficulties wouldn't it?

Nvidia wanted to pull the Dx standard towards optimization on their hardware. That did not happen and Nvidia went it's own way. There is a close consultation between the developers of API's and the Graphics industry. You mention MS/ARB specs and Nvidia with its Design issues in NV30. ATI who developed their DX9 products EARLIER than Nvidia and got it right. Therefore your argument that "companies under or overestimate build processes" is not based upon a lack of information from the API's. Nvidia got it wrong, I think their dominant position in the graphics market lead them to believe they could be the "tail that wagged the dog". Nobody wags Microsoft. Are you saying that ATI had greater access than Nvidia? I don't think so. But if so, would it not only further the argument that an open consultative process of design is even more necessary? Read Tom's Editorial on this at T.H.- he argues the same way.
-Benchmarking Games Conclusion

Another question then

-If Nvidia had designed for 24 bit precision instead of 16and 32 bit would we not have a different market in this place and time?

Problems with the system of design,development and implementation can be managed. What are the alternatives? An even more fragmented and divisive process? That would be in nobodies interest.

 

[DaveBaumann]

I was only talking about the 99% of games that did not have an option to select AF.

That didn't appear to be what you were saying at all:

Whether or not you select trilinear in game you aren't getting it.

If you think that is only older games, Halo doesn't have the option.

If thats what the developer feels then thats what the developer feels. Applying something through the control panel goes outside of the specification of the application and is then up to the user and the board that they purchase.

Not compiling the code for the market leader...? Why would a developer with lesser resources ignore the largest market?

Lets reapeat this one more time: ATI had the vast majority of the high end DX9 market and largest slice of the NVIDIA's DX9 market was a for a board that is/was primarily being treated as DX8, in other words they were not the market leader for their target DX9 base.

And the common theme here is that everyone is going to jump on the HL2 engine and noone would dare use Doom3 because of the fact that it doesn't use PS 2.0. I know you aren't saying it(though you have been supporting the widespread PS 2.0 penetration soon), but that is the illusion that many people are giving to people. HL2 is all that matters, Doom3 is DOA.

Well, I've not read eveyones comments here, however it doesn't seem that way - could it just be the way you are interpretting things? However, as I've said before I do question whether the D3 engine will see such widescale adoption as previous id engine due to the inherant nature of the lighting model not coping well with outdoor scenes. I want to see if D3 has any levels outdoors at all, and if so how it handles them.

Yes, you did say it was from the E3 build and expanded to state that is why HDR was not used as it had not been implemented at that point.

Well, I shuldn't have said build in that context - it was a build prior to the HDR rendering (as this was very new) and the benchmark featured the E3 demo levels. Chonologically it would seem that it had to have been a later build since as they had only worked on any FX board just prior to E3 the MixedMode couldn't have been implemented, whereas it was in the benchmark.

 

[Pete]

Spam, currently there are two (major) 3D APIs: Direct3D, which limits you to Windows platforms, and OpenGL, which allows for Windows, Linux, and Apple.

 

[spam]

Thanks Pete,
I edited my previous post accordingly

 

[BenSkywalker]

JB-

You pointed out that Halo did not show same delta.

I pointed out that Halo does not have a full precisions mode and uses partial precision for FX

PP for the NV3X under HL2 also showed the R9800 with a substantial lead while Halo has the 5900 ahead of the 9800.

John C has already said their IS an IQ difference in D3 between the two. Its gonna depend on the game I feel.

He said there was 'no discernable quality difference', it's in one of B3D's interviews. Of course in the theoretical sense there is, as there is between FP32 and FP24.

The general conclusion looking at TR, HL2, Halo, Shadermark, 3dmark, Rightmark,ect is that the R9800 is much faster running PS2.0 shaders.

And the one actual game that is out that people want to own at that list is as fast or faster on the 5900. That is what I can see with crystal clarity.

I am sure PS2.0 will be come more limiting factor. But you have to remember not every game will use them in the same manor. AQ3 for example only has one PS2.0 shader. HL2 had a ton of PS2.0 shaders.

When will PS 2.0 become more of a limiting factor? We have been hearing that this will happen 'soon' for about a year now and so far we have two titles shipping, one of them worth owning and that happens to be faster for the most part on nV hardware.

Radeon 9800 Pro, stock and Radeon Pro 256 MB (OVERCLOCKED 405/371: ), Catalyst 3.7 drivers (Cheat detection enabled / disabled...no difference)
GeForce 5900 Ultra, Det 52.xx, PS 2.0a, partial precision mode.

It ends up showing close to what I would expect in terms of performance rift. I understand there is a difference but....

Notice the how many test the FX can not even run. Shall we factor those in

Refrast doesn't handle some of those properly. If refrast can't do it, then no way should any board be expected to.

Does it make more sense to map your benchmark to the DX9 standard or map it to something non-standard?

PS2.0A is part of the standard.

Spam-

Departures from these API's(OpenGl and Direct3d) would only add to the difficulties wouldn't it?

Here's the problem, DX10 parts are being designed right now, DX10 won't be finalized for some time yet. Who will we end up saying went wrong? As of right now I'd have to say nVidia will be the odd man out as they are aren't the next XBox chip. What nV ends up having to do is take a guess as to where ATi is going to go and try to exceed that for their NV50. If they waited for the DX10 specs to be finalized prior to beginning work on their DX10 part they would be dead in the high end for years.

ATI who developed their DX9 products EARLIER than Nvidia and got it right.

Microsoft adopted ATi's submission for DX9 over nVidia's.

Are you saying that ATI had greater access than Nvidia?

Yes, and it isn't just me saying it. You can ask the others involved in this discussion, MS went with ATi's submission for DX9. The reverse happened for DX8, MS went with nVidia(which everyone knew was going to happen due to the XBox).

But if so, would it not only further the argument that an open consultative process of design is even more necessary?

You can run in to issues with this. The IHVs are not going to want to show all of their cards, they are not going to want to let the other know exactly what they are doing. For nVidia on this gen of parts, they decided to offer FP16 which is what Carmack has been asking for along with FP32 which had a preexisting standard(IEEE) and also is useful for their Quadro parts for use in preview rendering in Viz applications. Looking at their part I would wager that they were of the mind that game developers were going to be utilizing FP16 for their shaders until they made the leap to FP32(Sweeney has mentioned this is his intention as a side note, combined with Carmack's desire for FP16).

For the DX10 parts everyone will be FP32, it will allow combining the shaders units so you won't have to deal with seperate Vertex and Pixel shader units. It's easy in hindsight to say shoulda/coulda/woulda, but based on the information nV had I don't think their choice in terms of accuracy was a bad one.

If Nvidia had designed for 24 bit precision instead of 16and 32 bit would we not have a different market in this place and time?

It would have hurt their potential on the pro part end. If they had the same register limitation it actually would have hurt them in the gaming space as they would be forced to have partial precission in 12INT instead of FP16.

Problems with the system of design,development and implementation can be managed. What are the alternatives? An even more fragmented and divisive process? That would be in nobodies interest.

Things are actually a lot better now then they have been in the past. If you looked at specifications that existed when nV and ATi went in to the design phase nV is the only one that was following an existing standard. What's more, they also took the path with FP32 that ATi will be following with their DX10 parts. Neither the R3x0 or NV3X are fully DX9 compliant, neither of them are close actually, but the one thing we are left arguing about is their differences in PS performance the majority of the time. Look back to the days of 3dfx where they seemed utterly he!l bent on leaving out every feature they could, refusing to advance unless they could do so in a proprietary method the overwhelming majority of the time. Back when it was nV versus 3dfx in the gaming enthusiast market, nV did what they could to end up being closest to the DX spec. They never got it completely right back then either(prior to the XB deal and the NV2X line being the basis for it), they were simply the closest out of the big two. They are back to being 'close but not quite' there. What if we had the Rampage3 showing up with PS 1.1 and VS1.1 support but it had the next generation TBuffer3 effects? Really, with ATi and nVidia right now they both have a fairly close view on where the industry is headed and even without standards set in advance they can come out with parts that are quite competitive.

Would it be advantageous for the specs to be out early? Maybe, but what happens if MS guesses wrong? With the NV30 it is easy to blame nV for being too agressive with their expectations on build process, but if MS sets the standards and date for a new generation of parts(due to DX availability) and they have requirements that are either far too simplistic or complex for the build process available you end up with an entire multi billion dollar industry SOL.

Dave-

That didn't appear to be what you were saying at all:

That's what the quoting gets us, the conversation was in the context of 'cheating' when enabling AF.

If thats what the developer feels then thats what the developer feels. Applying something through the control panel goes outside of the specification of the application and is then up to the user and the board that they purchase.

With the board they purchased they have quality mode sliders that are supposed to enable AF and full trilinear, you are not getting this with either company right now if you enable it in the CP. I consider both of these implementations a hack, however if one of them is cheating, then both of them are.

Lets reapeat this one more time: ATI had the vast majority of the high end DX9 market and largest slice of the NVIDIA's DX9 market was a for a board that is/was primarily being treated as DX8, in other words they were not the market leader for their target DX9 base.

'High end DX9 parts' is a miniscule market. If a game is coming from a smaller developer they had better target it so it can run on a 5200 if they don't want to remain a small(or dead) developer.

Well, I've not read eveyones comments here, however it doesn't seem that way - could it just be the way you are interpretting things?

Absolutely not. I'm not talking about this thread, I'm talking about on these forums. There are forum regulars in this discussion, I don't think it leaves much to try and figure out.

However, as I've said before I do question whether the D3 engine will see such widescale adoption as previous id engine due to the inherant nature of the lighting model not coping well with outdoor scenes. I want to see if D3 has any levels outdoors at all, and if so how it handles them.

There are outdoor areas in Doom3. Since you don't have a helmet in Doom3, you have to run around and grab air canisters while you are outdoors. The biggest 'problem' with Doom3's outdoor engine is that you lose a lot of the dramatic effect of the lighting as the infinite light is too strong(guess you have been following HL2 closer then D3 ). This has actually been discussed at B3D IIRC(though I may be recalling wrong I know I've seen it discussed in forums somewhere and I doubt it was here, id talked about this issue explicitly in an interview in one of the print publications, can't recall which one now but if I remember I'll look through my last few months worth of PC gaming mags and tell you which article it is in).

 

[Pete]

Originally posted by: Genx87
Does it make more sense to map your benchmark to the DX9 standard or map it to something non-standard?

Standard or not it would make a lot of sense to map it to hardware that is the majority of the market.

I'm pretty sure ATi's DX9 line owns both titles at the moment. (As 5200 and even 5600 speeds show, neither are worth considering as DX9 parts. I'll be kind to the 5600, but I can't honestly include the 5200 as a DX9 part.)

 

[BFG10K]

Ben:

You think they managed to get them up an running that quickly? That seems awful hard to believe to me.

It also seems unlikely that they managed to whip up no less that eight hardcoded cheats into 3DMark03 within the space of one driver release. But they did it, nevertheless.

JK2 is CPU limited anyway, but that would be a valid point for the older boards that would be vid limited.

Lackluster image quality is valid in any situation, regardless of the rendering mechanics of the game.

Not neccessarily. Tell me how the filtering optimisations would ever break a game?

A filtering "optimisation" isn't what I had in mind, I was talking about things like shader subsitution and similar. What happens if the game completely changes the shader? Or if it removes it altogether?

Look at a R2x0 board running AF, it is horrendous.

It's far better than straight bilinear/trilinear that's for sure. Also the performance hit is extremely minimal.

Besides, if you don't like it you can turn it off and actually gain a little speed too. If your turn off nVidia's DXT1 you lose speed.

That's why we have apps like RivaTuner and RageTweaker.

Right, now show me the standard registry key that nVidia provided to disable their application detection mechanisms and allow reviewers to use non-cheating driver paths.

You get trilinear on one mip level, and that's it. All the rest are straight bilinear, not even a brilinear hack. Have you not noticed this blatant 'cheating' yet?

That's correct. But the setting performs this universally regardless of API or application, unlike nVidia's setting which chooses to do absolutely nothing when it suits it. The setting also changes its mind between driver versions.

Also show me where ATi's screenshots are showing images that aren't actually rendered like nVidia's in order to hide their "optimisations".

Let's see how your R300 runs the FX compiled code when it is the only option.

It can't be any worse than the FX's inability to run certain code at all.

IPC? Please, let's keep it reasonable here, no need to go Mac user.

No, a Mac user is one who defends inferior hardware even in the face of overwhelming evidence and opposition that says otherwise.

And I'm sure you would be real happy with code that would not run at all on your board that was optimized for the FX.

And Glide doesn't run on an FX. So what? Are we going back to using proprietry code as examples again? For straight generic DirectX code the R3xx wins hands down and that's what developers are going to go for. Unlike Carmack, Sweeney and Gabe the smaller developers don't have the time or money to waste time making crutches for hardware that'll still be crippled at the end of their excercise.

They were trying to convince developers and people like you that there was a much larger performance gap then there actually is.

So I'll ask again: does a Microsoft compiled full precision path beat the mixed mode path? If doesn't stop this nonsensical argument and if it does, show me evidence.

TRAoD runs on the PS2's DX5 level hardware, you sure you want to use that line? OK, we'll go with it.

But we know it has PS 2.0 effects on the PC port because the devs said so. I want to see the evidence you mentioned about Halo doing the same thing.

But according to you TRAoD is DX5, right?

I'm not going to proceed down this ridiculous path except to point out the half a dozen or so PS 2.0 benchmarking programs that show the same hardware disparity between the NV3x and the R3xx.

 

[spam]

Ben,

The history of Direct 3d development is not definitive. Your review of recent history needs to be substantiated with some more corrobration. For my part I would say that Microsoft put forward the proposed standard which ATI accepted and NV rejected. At this juncture, it would be best to corroborate our statements with supportive evidence. I think my account is the more common recounting of events. Could you provide some evidence to support yours? I will try with mine.

 

[Genx87]

I'm pretty sure ATi's DX9 line owns both titles at the moment. (As 5200 and even 5600 speeds show, neither are worth considering as DX9 parts. I'll be kind to the 5600, but I can't honestly include the 5200 as a DX9 part.)

You can think whatever your little heart wants. Nvidia had 56% of the standalone market in the last quarter. ATi was a less siginificant 26-27%.

 

[BenSkywalker]

It's far better than straight bilinear/trilinear that's for sure.

Hehe, try actually using it. It has hideous level of aliasing.

Right, now show me the standard registry key that nVidia provided to disable their application detection mechanisms and allow reviewers to use non-cheating driver paths.

So far you have managed to say that 3DMark2K3 has cheats and that is the only thing. Your big problem is that 3DM2K3's old build with old drivers had cheat issues? You launch in to bashing nV at the drop of hat for that singular issue?

That's correct. But the setting performs this universally regardless of API or application, unlike nVidia's setting which chooses to do absolutely nothing when it suits it. The setting also changes its mind between driver versions.

So they are by your definition cheating, or nVidia is not. Either ATi and nV are both cheating or both of them aren't, take your pick.

Also show me where ATi's screenshots are showing images that aren't actually rendered like nVidia's in order to hide their "optimisations".

???? What are you talking about? Are you talking about ATi not rendering certain effects properly as was recently discussed at the counterpart to ShaderDay, or something else?

It can't be any worse than the FX's inability to run certain code at all.

Your board wouldn't be able to run it.

No, a Mac user is one who defends inferior hardware even in the face of overwhelming evidence and opposition that says otherwise.

IPC when comparing things that don't have comparable clock speed is pretty much the defining characteristic of a Mac user. When using IPC as a reason why something is superior to a competitor. Explaining performance differences related to IPC is one thing, using that as an edge is very much Mac user.

And Glide doesn't run on an FX. So what? Are we going back to using proprietry code as examples again?

Not even close. I'm talking about using the DirectX compilers. You compile the code with ATi's targetted compiler and it runs on a FX, do the same with the FX compiler and see what you get.

For straight generic DirectX code the R3xx wins hands down and that's what developers are going to go for.

If that was true then it would be a good point. It isn't and it isn't. You have to add a rather extensive list of qualifiers for that to be true, your generic statement isn't close to valid.

Unlike Carmack, Sweeney and Gabe the smaller developers don't have the time or money to waste time making crutches for hardware that'll still be crippled at the end of their excercise.

The reality is that smaller developers can't afford to support the utterly miniscule marketshare of high end DX9 parts as a focal point.

So I'll ask again: does a Microsoft compiled full precision path beat the mixed mode path? If doesn't stop this nonsensical argument and if it does, show me evidence.

How is it nonsenical to take issue with your repeated false statements that nV is destroyed in DX9 performance.

But we know it has PS 2.0 effects on the PC port because the devs said so. I want to see the evidence you mentioned about Halo doing the same thing.

Pixel shaders 2.0 (DirectX 9.0)
In this code path, you are making absolutely no compromises on the visual quality of the game. You are seeing everything as best as possible, as engineered by our team. All the effects are in their most demanding form (as complex of a calculation as necessary to generate the best visual result possible).

Having said this, for many simple effects, even if you are running PS2.0, the game will automatically use a 1.4 or 1.1 shader because the visual result is exactly the same.

Link. This information has been posted hundreds of times around the web. This being a game people actually want to play I guess you just missed it, keep having fun with TRAoD- have you beat that yet btw? I know how highly you think of the game, it being the definitive point of DX9 based on your comments.

I'm not going to proceed down this ridiculous path except to point out the half a dozen or so PS 2.0 benchmarking programs that show the same hardware disparity between the NV3x and the R3xx.

Like ShaderMark? The bench that has multiple tests that RefRast can't render properly and has another test built around a bug in DirectX that MS has stated they intend to remove? Or like 3DMark2K3 which by design was built to render as slow as possible? You can't be talking about the one game out using DX9 shaders that people want to play, as that is in direct contradiction to your repeated insinuations that nV is going to have major issues with all DX9 titles.