Anand's 9800XT and FX5950 review, part 2

[Rogodin2]

"So, you don't think it's misleading to advertise the card as DX9 compliant even though it's going to run in DX8 mode?

Misleading? Depends..........................

If you are buying an 80 dollar card that is being advertised to run a DX9 class game at the highest resolution at high speeds then yes. If you are buying a DX9 (compliant) card because it is (compliant) then no.

Also, I'd like to point out the inconsistency of your statements:

Here let me help you get this correct......

Clueless n00b who buys an OEM machine or the cheapest part they can find. They get a 5200 which is a DX9 compliant card. When they plop it into their machine and run a game with DX9 shaders the game will probably default to a DX8 path and the person would never know the difference.

and

Again nobody who is buying an 80 dollar video card is going to expect it to run the newest greatest games as 100FPS with all the goodies on. They expect it to perform well at lower end resolutions and settings. There wont be mobs of irate 5200 card owners because those people know what they are getting into when they plop down 80 bucks for one.

First you say the people who buy these cards don't know what to expect and then you say they really do."


This is all conjecture and drivel.

You are making statements that are incapabale of being validated and they you use them to support your claim??????

You might want to brush up on your logic.

rogo

 

[BenSkywalker]

If you shift to a different compiler or even a different version of the same compiler your code isn't automatically gauranteed to work unaltered, especially if your original code was built to work around compiler issues with your previous compiler.

It isn't like we are talking about complex compilers on either end. The one aimed at ATi is extrodinarily simplistic compared to any x86 compiler of the last couple decades. The functionality of PS 2.0 is still very, very simple compared to any CPU compiler. There could be some bugs with some of Valve's code(ATi has demos that use some DX9 bugs as 'features'), I don't consider the odds of that to be much above a rounding error. If there were, they could easily have stated as much.

If you shift to a different compiler or even a different version of the same compiler your code isn't automatically gauranteed to work unaltered, especially if your original code was built to work around compiler issues with your previous compiler.

Link?

I don't want to publicly promote their behaviour.

Given Doom III only uses a few simple 1.x shaders I wouldn't expect OOE to help it much, if at all.

It's doing what can be done using 1.x level shaders, but it isn't doing them in that manner on NV3X or R3X0 hardware. It is using the PS2.0 level functionality to collapse the amount of passes required. Their driver level optimizations should prove to offer a decent improvement, although I wouldn't expect quite the same level as we have seen in some other titles as we are talking about Carmack's code.

That sounds like a reasonable theory but I have problems giving nVidia the benefit of the doubt after everything we've seen from them.

Most of what we have seen involves assumptions about what they were doing in regards to their shaders. They were busted in 3DM2K3, other then that we have had a large portion of the community jumping to conclusions.

The NV3x has performance hits when running 1.x shaders too, just not as much as it does with version 2.0.

They are relatively close to even with ATi running PS1.x, and even if they weren't MP2 isn't a pixel shader heavy game.

Yes but the 9800XT came out what, two weeks later? Valve's plan must've been to benchmark current cards at the time and then allow reviewers to again use the HL2 benchmarks when ATi's XT boards arrived. Given ATi is Valve's bundling partner there's no doubt in my mind that the benchmarks were timed to coincide with the 9800XT's release.

We knew well in advance when the launch date for the XT was and yet Valve pushed out the bench number ahead of that by a decent margin.

Personally I believe that if a program requests something it should always get it unless the hardware isn't capable or the user has overidden the request.

But if you force enable quality AF and select trilinear filtering in game(twice asking for trilinear) you aren't getting it with ATi either. ATi is doing the same thing, actually with more frequency then nVidia when you factor in OpenGL.

 

[Genx87]

This is all conjecture and drivel.

You are making statements that are incapabale of being validated and they you use them to support your claim??????

You might want to brush up on your logic.


Well since we are talking about market issues, peoples personal preferences, and percieved performance. I think it is pretty obvious there are no "hard" facts to back up any of our claims.

Um,

Duh!

 

[BFG10K]

Ben:

I don't want to publicly promote their behaviour.

I hope you don't mind but I'd like to link the article and respond to your PM in this thread because I don't believe there's anything wrong with doing so.

HalfLife 2 beta benchmarks.

Anyway, in response your PM:

with AA+AF they are actually quite close.

Where should I be looking? From those graphs the Radeon cards have a substantial lead across the board. In fact the 9800XT is often more than double the speed of the 5950, and that's with the Detonator 5x.xx drivers which include realtime shader reordering. From those results it appears the situation is even worse for nVidia than it was before unless the mixed mode path isn't being used.

Some of the 5600/5700 & 9600 scores are close but that's when the cards are both unplayable and crippled from their lack of memory bandwidth due to the settings they're being requested to do, settings which do not indicate shader performance.

Also if you read the commentary the FXs have some issues with image quality yet again. This is a recurring problem in all popular benchmarked games and this is why it really makes it difficult to believe that are simply one-off bugs.

It is using the PS2.0 level functionality to collapse the amount of passes required.

My understanding was that it was using PS1.4 shaders to perform PS 1.1-1.3 operations with a lower number of passes. I wasn't aware that Doom III was doing any PS 2.0 shading at all.

They are relatively close to even with ATi running PS1.x,

I'm not so sure about that. It appears that even NV25 hardware is faster at running PS 1.x than the NV3x is and it's widely known that the R3xx boards have absolutely no performance problems with it at all. In fact even the R200 had extremely strong PS 1.x performance back in its day, higher than the NV25 at times.

We knew well in advance when the launch date for the XT was and yet Valve pushed out the bench number ahead of that by a decent margin.

And I believe it was timed for the XT's launch.

But if you force enable quality AF and select trilinear filtering in game(twice asking for trilinear) you aren't getting it with ATi either.

But the game isn't requesting AF. If it was that would be a cheat.

 

[Rogodin2]

"Well since we are talking about market issues, peoples personal preferences, and percieved performance. I think it is pretty obvious there are no "hard" facts to back up any of our claims.

Um,

Duh"

Convoluted post but I think I catch your drift-don't post your inanities ESPECIALLY if you KNOW they're inane in the first place.

rogo

 

[BenSkywalker]

Where should I be looking? From those graphs the Radeon cards have a substantial lead across the board. In fact the 9800XT is often more than double the speed of the 5950, and that's with the Detonator 5x.xx drivers which include realtime shader reordering. From those results it appears the situation is even worse for nVidia than it was before unless the mixed mode path isn't being used.

The top bench, ixbt07, the 5950 and 9800XT are within 6.9FPS of each other with AA and AF on, and that is at the lowest resolution they tested at(still within playable framerates). On the less demanding bench ATi still does have a decent lead, but it isn't the 70% we saw in the past. I'm not entirely sure they are using MM in the tests as of this point, if they are then I think the impact on IQ using MM over pure was overstated quite a bit.

Some of the 5600/5700 & 9600 scores are close but that's when the cards are both unplayable and crippled from their lack of memory bandwidth due to the settings they're being requested to do, settings which do not indicate shader performance.

They are a lot closer then they were.

Also if you read the commentary the FXs have some issues with image quality yet again. This is a recurring problem in all popular benchmarked games and this is why it really makes it difficult to believe that are simply one-off bugs.

You mean them missing the shader on one mip level in one screenshot?

My understanding was that it was using PS1.4 shaders to perform PS 1.1-1.3 operations with a lower number of passes. I wasn't aware that Doom III was doing any PS 2.0 shading at all.

You could say it isn't PS2.0 as it is working under OGL, but it uses functionality that the DX8.1 class parts lack to collapse the rendering in to a single pass(not to mention using FP16).

I'm not so sure about that. It appears that even NV25 hardware is faster at running PS 1.x than the NV3x is and it's widely known that the R3xx boards have absolutely no performance problems with it at all.

Link. Are you talking about the low end/older NV3X parts? The NV35/36 handily whip the NV25 in DX8 shaders(around 30% outside of 640x480 for the NV36, the NV35 is better still).

And I believe it was timed for the XT's launch.

Where are the XT benches? If it was timed for the XT launch, why didn't we see any XT numbers for their unveiling of the bench? Why have the big PR event when they did?

But the game isn't requesting AF. If it was that would be a cheat.

So it is a cheat to you. The game is requesting trilinear and not getting it. I just want to make sure I had it clear.

 

[Genx87]

Convoluted post but I think I catch your drift-don't post your inanities ESPECIALLY if you KNOW they're inane in the first place.

Unfortunately I dont think you "really" get my post.

But continue.........................

 

[BFG10K]

The top bench, ixbt07, the 5950 and 9800XT are within 6.9FPS of each other with AA and AF on,

That's not a shader limited situation though, it's when memory bandwidth limitation is kicking in. You can see this because of the large boost the Radeon cards get when disabling AA but the FX line largely retains the same performance level. Since we already know the non-AA benches aren't as memory bandwidth intensive and we know that the benchmarks aren't CPU limited that leaves only one other option - shaders.

On the less demanding bench ATi still does have a decent lead, but it isn't the 70% we saw in the past.

When it counts the 9800XT is more than double the speed of the 5950. That is a big difference as the two cards aren't even in the same league in terms of shader performance.

They are a lot closer then they were.

Yes but they're meaningless.

You mean them missing the shader on one mip level in one screenshot?

From the article:

In case of the FX 59xx at the distance of two MIP levels there's blurred strip on the water that removes reflections.

but it uses functionality that the DX8.1 class parts lack to collapse the rendering in to a single pass(not to mention using FP16).

DX 8.1 = PS 1.4 and that's why I speculated that the instruction reordering won't help it that much.

Link. Are you talking about the low end/older NV3X parts?

The link is dead as of this moment. My comment was based on Ti4600 cards beating most FXs in HL2 benchmarks when running the DirectX 8.x path.

Where are the XT benches?

They arrived a few weeks later.

Why have the big PR event when they did?

To get everyone used to the current situation by allowing the current benchmark results to sink in and then allow ATi to deliver something better shortly afterward.

I just want to make sure I had it clear.

nVidia promotes trilinear AF as a feature on their FX series. Now explain to me how a developer or user goes about enabling such a feature under Direct3D.

 

[BenSkywalker]

Since we already know the non-AA benches aren't as memory bandwidth intensive and we know that the benchmarks aren't CPU limited that leaves only one other option - shaders.

Not exactly, there is AF combined with shaders. I hope nVidia 'fixes' this soon and does the same level of filtering as ATi does on shaders when AF is enabled.

When it counts the 9800XT is more than double the speed of the 5950. That is a big difference as the two cards aren't even in the same league in terms of shader performance.

Missed that one, where was that at?

From the article:

In case of the FX 59xx at the distance of two MIP levels there's blurred strip on the water that removes reflections.

That's what I was asking if you were talking about or not, the missing shader on the one mip level in that one screenshot. Just wanted to make sure I wasn't missing anything.

DX 8.1 = PS 1.4 and that's why I speculated that the instruction reordering won't help it that much.

Huh? By using the features that the DX8.1 parts lack they are collapsing multiple passes in to a single, driver rescheduling should help here(actually, id already said it did). Saying PS 1.4 doesn't quite work as we are talking about OGL, differing levels of functionality nite quite comparable to DX(nv exposes certain things that don't work under D3D).

The link is dead as of this moment. My comment was based on Ti4600 cards beating most FXs in HL2 benchmarks when running the DirectX 8.x path.

You were talking about the FXs being slow running DX8 level shaders, comparing the line up overall the FXs are significantly faster running DX8 shaders then the NV25.

They arrived a few weeks later.

Where? Where are the HL2 XT benches that aren't from the leaked build?

To get everyone used to the current situation by allowing the current benchmark results to sink in and then allow ATi to deliver something better shortly afterward.

That didn't happen.

nVidia promotes trilinear AF as a feature on their FX series. Now explain to me how a developer or user goes about enabling such a feature under Direct3D.

Want to try it another way? I'll list games that full trilinear and AF work on for nV right now, you do the same for ATi both using their latest drivers. nV outnumbers them by a significant margin(nV works under all OGL, ATi only works on a small handful of each API). That is the reality of the situation. If you insist on considering nV cheating, ATi is in the same boat.

 

[BFG10K]

Not exactly, there is AF combined with shaders.

AF is not responsible for the performance degradation. There were non-AF HL2 benchmarks out before that showed the same performance delta.

Missed that one, where was that at?

The Coast demo with the high-end hardware. The green brick bars are basically half the size of the red brick bars and ATi is running a higher level of AF too.

Huh? By using the features that the DX8.1 parts lack they are collapsing multiple passes in to a single, driver rescheduling should help here

Then why do FX cards run faster under the DirectX 8.1 path than they do under the DirectX 9 path? Surely the reduced passes should be increasing the speed when running PS 2.0 code?

You were talking about the FXs being slow running DX8 level shaders, comparing the line up overall the FXs are significantly faster running DX8 shaders then the NV25.

NV30 based cards are not faster than NV25 cards except maybe the 5800 Ultra if it has enough brute force. NV35 cards probably are though.

Where are the HL2 XT benches that aren't from the leaked build?

I'm having difficulty finding them right now. When I find them I'll give a link.

I'll list games that full trilinear and AF work on for nV right now, you do the same for ATi both using their latest drivers.

You aren't looking at the stuation and you're simply going off on a tangent.

The situation is that nVidia is denying all developers and users the ability to do trilinear AF in Direct3D regardless of what is requested simply to boost performance. How many games it effects is completely irrelevant. Do you think Microsoft would make any difference to handing out WHQL certification if a driver that didn't follow a hypothetical spec caused an issue in only one game as opposed to affecting 100 games?

If you can't see that as a cheat then I really am baffled.

If you insist on considering nV cheating, ATi is in the same boat.

I'd call ATi's scheme a partial cheat but I wouldn't call it as being on the same level as nVidia's.

 

[bluemax]

I noticed both cards seem to do extremely well and are both priced relatively close together.

My next card will be some sort of 9500/9600/9700/800 Variant.

Why? Simple... only the latest Radeon cards have the capability and software to enhance DIVX playback.
That's important to me.

I like the fact ATI has tried very hard to make an all-around computing experience right down to video playback (always has) as well as games.

nVidia seems to be about games and nothing but...... I do more than just play games.

 

[BenSkywalker]

AF is not responsible for the performance degradation. There were non-AF HL2 benchmarks out before that showed the same performance delta.

It is responsible for a disproportionate performance degredation. There are issues with AF and pixel shaders, if you have picked up Halo yet go in to one of the levels covered with shaders and take a shot at both 16x Q AF and no AF and compare the results. Check out Kyle's Halo IQ comparison where they have both the 5950 and 9800 screenshots up and compare them. nVidia's AF is screwed up when filtering shaders, ATi doesn't seem to be doing anything at all. Given what AF does, I'm not bashing ATi at all on this one, I don't think it is very viable to filter them properly.

The Coast demo with the high-end hardware. The green brick bars are basically half the size of the red brick bars and ATi is running a higher level of AF too.

When it counts the 9800XT is more than double the speed of the 5950.

I'm putting those two together as there aren't any benches that show the board twice as fast, that's why I was asking which bench it was.

Then why do FX cards run faster under the DirectX 8.1 path than they do under the DirectX 9 path? Surely the reduced passes should be increasing the speed when running PS 2.0 code?

We were talking about Doom3, are we still on the same page here? If you are talking about HL2 then that's something else entirely.

NV30 based cards are not faster than NV25 cards except maybe the 5800 Ultra if it has enough brute force. NV35 cards probably are though.

Let me check on that..... nope. The link has the NV30 clocked at Ultra speeds and @400/400 but the difference exceeds twice as fast versus a Ti4600 even @400/400(125% at 1600x1200, 162% for the Ultra). I think you are thinking the NV3X parts are a lot weaker then they actually are

The situation is that nVidia is denying all developers and users the ability to do trilinear AF in Direct3D regardless of what is requested simply to boost performance.

ATi is denying you the right to select trilinear + AF in almost every single game that you own no matter what you request in order to boost performance. I was not going off on any tangent, I was talking directly about who is 'cheating'. ATi refuses to filter properly no matter what you request in the overwhelming majority of games. In fact, there are only a very small handful of titles where you can run full trilinear and AF with ATi's parts even if you request it, less then there are for nVidia in fact. If one of them is cheating, both of them are and by looking at the amount of titles it impacts, ATi has the larger number.

I'd call ATi's scheme a partial cheat but I wouldn't call it as being on the same level as nVidia's.

I agree with you there, nVidia's effects less games then ATi's

I'm having difficulty finding them right now. When I find them I'll give a link.

I've looked all over, with the exception of the above linked article, I haven't found any. If there intention was to see it benched on the 9800XT then we certainly would have seen said benches on AT, H and Tom's at least.

 

[BFG10K]

It is responsible for a disproportionate performance degredation.

Like I said before, check the initial HL2 benchmarks without AF or AA and you'll see the same pattern as you're seeing here. When you take away the memory bandwidth limitation the R300s are generally twice as fast as NV3x cards in the full precision DirectX 9 path.

I'm putting those two together as there aren't any benches that show the board twice as fast, that's why I was asking which bench it was.

27.1 vs 51.8.
35 vs 67.

Not exactly double the speed but close enough. Certainly much closer and reaslitic than your 0.1 FPS example you used to show the "superiority" of nVidia hardware in Halo (never mind the fact that you totally ignored the AA/AF lead that ATi took).

We were talking about Doom3, are we still on the same page here?

What difference does it make? You're making the claim that PS 2.0 hardware is automatically faster than PS 1.x hardware because it can render in fewer passes. I'm saying it isn't because there are other factors involved.

ATi is denying you the right to select trilinear + AF in almost every single game that you own no matter what you request in order to boost performance.

No they aren't. They are denying trilinear past the first mip-map, unlike nVidia who is doing bilinear on everything. Score 1 to ATi.
Also ATi does full trilinear AF when the app requests it, unlike nVidia who do not. Score 2 to ATi.

I agree, ATi is marginally cheating but nVidia's cheat is worse and isn't marginal by any means because they're employing blanket tactics to make sure it's always in effect. ATi's method is doing it in a roundabout way since technically the app isn't requesting AF.

I agree with you there, nVidia's effects less games then ATi's

So then by your logic nVidia's clip plane cheat in 3DMark03 is less of a cheat than their bilinear AF since it only impacts one title vs the BiAF which impacts all Direct3D titles?

Ironically using your previous statements you don't even consider nVidia's forced bilinear AF being a cheat, yet at the same time you're attempting to explain that ATi is cheating more than nVidia because their actions impact more games than nVidia's do. So which is it? Are we using the number of games to measure what a cheat is or aren't we?

I'm not and I never was. A cheat is a cheat for its own reasons and the number of games it impacts is irrelevant.

I've looked all over, with the exception of the above linked article, I haven't found any.

I saw at least two that issued addendum scores to their first HL2 benchmarks but they were linked from other forums and I'm having a hard time finding those threads again. I remember the XT was about 10-15% faster than the 9800. I'll keep looking though.

If there intention was to see it benched on the 9800XT then we certainly would have seen said benches on AT, H and Tom's at least.

I agree, that is strange since AFAIK Valve didn't "expire" the public benchmark so I can't see any reason why sites didn't publish newer benchmarks.