Ati promises Doom3 performance Boost!

[Keysplayr]

Originally posted by: Insomniak

Originally posted by: keysplayr2003

If this holds true, shouldn't even the 12 pipe 6800 be double the performance of the X800pro in DoomIII? Because I dont see that happening. A 16 pipe GT or Ultra doesn't even double the performance.

It has nothing to do with performance doubling - only doubling the number of pixels rendered per clock. This results in a hefty boost to performance, but games aren't rendered in pixels - they're rendered in frames, which pixels go in to. Having faster pixel rendering will definitely help FPS, but FPS increase and pixel render speed are not directly correlated.

Ok lemme see if I get this right. Card "A" renders 1/2 as many pixels as card "B" in the same amount of time. Pixels go into frames and the frame can only be rendered when its filled with pixels. Therefore it would take card "A" twice as long to render a frame then card "B". But you say fps increase and pixel render speed are not directly correlated. You see where I'm confused here right? So how does it work, really. Unless you explain what you actually mean, it looks like your contradicting yourself in the same sentence. So have at it.

 

[reallyscrued]

quote : That's bullsh*t

The original Unreal Engine was GLIDE ONLY. Direct3D and OpenGL support was added in patches. Unreal Tournament shifted the engine's rendering focus over to Direct3D, but it was still basically a patch-together hack. Unreal Engine 2 was built from the ground up to take advantage of Direct3D, but it still supports OpenGL rendering.

Unreal Engine 3 will be built with DX9 targeted as the minimum spec, and will support OpenGL rendering.

END QUOTE

whoops, looks like i didnt do my homework. sry man, the unreal i got was coupled with GOTY edition of UT so i got the patched DX version of Unreal and neva knew bout the glide only one back in the day. but so far, i dont think, in all the games based on any unreal engine, gets better performance by picking OpenGL than DirectX. (ive only tested this on ATI cards though.)

Sry quote button dont work....dunno why....ill check it out in forum issues.

 

[gururu]

Originally posted by: reallyscrued

whoops, looks like i didnt do my homework. sry man, the unreal i got was coupled with GOTY edition of UT so i got the patched DX version of Unreal and neva knew bout the glide only one back in the day. but so far, i dont think, in all the games based on any unreal engine, gets better performance by picking OpenGL than DirectX. (ive only tested this on ATI cards though.)

Sry quote button dont work....dunno why....ill check it out in forum issues.

you may have not had the details correct, but for all practical purposes you weren't too far off. Unreal ran best in GLIDE and the UT's have always performed best (IME) in direct3d.
that they are compatible with OpenGL doesn't imply much.

 

[Insomniak]

Originally posted by: keysplayr2003
Ok lemme see if I get this right. Card "A" renders 1/2 as many pixels as card "B" in the same amount of time.

I don't know where the hell you got this from. I never said this, nor did anyone else in this thread that I can see.

Originally posted by: keysplayr2003
Pixels go into frames and the frame can only be rendered when its filled with pixels.

More appropriately, when all the pixels needed to complete the frame have been properly rendered. Different pixel types render different ways on different architectures using different drivers. I could put it differently, but that would be...too different.

Originally posted by: keysplayr2003
Therefore it would take card "A" twice as long to render a frame then card "B".

Here you are back at this two card example again. When did we start talking about cards?

Originally posted by: keysplayr2003
But you say fps increase and pixel render speed are not directly correlated.
You see where I'm confused here right?

Not really. There's a hell of a lot more to rendering a frame than just how fast your pixels come out of the pipe. Familiar with the differences in ARB, ARB2, NV10, NV20 render paths? If not, I'm not going to bother explaining them here. Go google.

Originally posted by: keysplayr2003
it looks like your contradicting yourself in the same sentence.

I don't see how that's the case, considering half the stuff you mentioned up here wasn't said by anyone.





Let me try this again:


Basically, a Doom 3 shadow is an untextured, unlit polygon (for soft, alpha blended onto whatever's behind it). Therefore, in simple terms, the pixel used to render them doesn't have to take up any TMUs when going through a pixel pipeline. It goes pretty much from triangle setup to Z culling to frame buffer.

This is true for both NV4x and R4xx cores. In theory, this means the number of pipelinesx2 = maximum pixel output per clock in Doom 3 PROVIDED THAT:

1) 50% of the pixels being rendered that clock are stencil shadows, and 50% are not.
2) The stencil shadows being rendered are not alpha blended.
3) The non-stencil shadow pixels are able to be rendered in one pass.
4) The drivers/firmware pushing the architectures are equally efficient in their organization and queueing (sp?) of instructions.
5) The fetching of texture data from the memory occurs at equal rates.
6) etc etc etc.


This is why you sometimes hear of NV30 referred to as 8x0 and NV40 referred to as 32x0. 16 pixels per clock with 1 texture per pixel, or 32 pixels per clock with no textures.

Suffice it to say, this situation never happens. But assuming all other things being equal, the number of pipelines a card has directly affects how many pixels it renders per clock in Doom 3.

So my original statement, which is that a card with 16 pipes should outperform a card with 12 pipes, holds true.

The reason performance isn't doubled just by having double the number of pipelines is simple. See 1 - 5 above. Note that the cores of the X800pro and 6800GT (to use two arbitrary examples) are clocked at different speeds. Note that ATi's software suite is obviously not as efficient at driving OpenGL rendering as Nvidia's is. Note 500 other x-factors.

The fact that, all other things being equal, double pipelines = double performance doesn't mean much when in the real world double pipelines does not equal double performance. This is because in the real world, all other things are not equal. However, despite that, having extra pipelines WILL give a given card a significant performance advantage when using the Doom 3 lighting model.

Can you sleep at night now?

 

[Pete]

Originally posted by: Insomniak
You're not familiar with the situation at hand here, so it's no surprise that you're talking pretty much out your butt here.

With the Doom 3 tech, pipelines pretty much ARE the end all be all. Do a little reading on how the lighting in the game is rendered and you'll see why. Basically, pipelines get a two for one deal when doing stencil shadowing - effectively allowing NV40 cores with 16 pipelines render up to 32 pixels per clock.

Even if ATi's 12 pipe X800pros were functioning as efficiently as NV's cards (which they don't, mainly because ATi's OpenGL support is poop), you're still looking at 24 pixels, tops.

Stencil shadows put a premium on the number of pipelines you're packing when it comes to performance.

Speaking of talking out the wrong end....

The GF 5 and 6 series do not output twice the pixels per clock with D3's stencil shadows, they just perform twice the z-/stencil-operations per clock (because--IIRC--you can use the color ROP for an additional z-op). The number of color pixels a GPU can output to the framebuffer is still limited by the number of pipelines (each with their attendant color ROP) it has (which for the 5800/5900 is four; the 6800, 12; and the 6800GT/U, 16).

Technically, though, stencil shadows put a premium on the number of z- or stencil-ops you can perform per clock, not the number of pipes (although the two are still linked, I think*). You can see this because a 6800's framerate is nowhere near its theoretical pixel rate (A pixel[ pipe]s/clock * B clocks/s = C GP/s), let alone its theoretical "zixel" rate (X pixel pipes/clock * 2 zixels/pixel pipe * Y clocks/s = Z Mzixels/s). This, in addition to the anticipated explosion of pixel shader effects, is why nV said they're seeing a 3:1 shader:texture ratio in new games. GPUs are starting to have to do a lot more work per pixel before they can output a finished product to your monitor, so they'll need a lot more back-end grunt than front-end flash with newer games that touch every pixel with a bunch of shaders, AFAIK.

This is a technical point, though, as I doubt having to touch the majority of pixels in a scene with lights will result in pixels per clock, but rather clocks per pixel render times.

As to the OP, I'm fairly confident that ATi won't catch up to nVidia in Doom 3, simply due to architectural limits. They should get closer (the X800P in particular looks way too slow), but the raw numbers are just stacked in nV's favor.

* This may not be the case with the 6800. A B3D thread noted that the 12-pipe 6800 may have a quad disabled, but kept all its ROPs (raster operations units) due to the fact that they appear to be linked to the quads via a fragment crossbar switch, not physically part of a pixel pipe (much like the memory). Thus, a 12-pipe 6800 it may have the same z- or stencil-op computational ability per clock as a 16-pipe 6800GT/U. I guess we'll see for sure when in-depth D3 benchmarks hit (probably courtesy of B3D).

Edit: OK, I can see how simplicity may be more understandable than what I posted above.

 

[Insomniak]

Originally posted by: Pete
Speaking of talking out the wrong end....

The GF 5 and 6 series do not output twice the pixels per clock with z-only operations, they just perform twice the operations per clock. The number of pixels they can output to the framebuffer is still limited by the number of pipelines they have (which for the 5800/5900 was four; the 6800, 12; and the 6800GT/U, 16). Technically, stencil shadows put a premium on the number of z- or stencil-ops you can perform per clock, not the number of pipes (although the two are linked). This, in addition to the anticipated explosion of pixel shader effects, is why nV said they're seeing a 3:1 shader:texture ratio in new games. GPUs are starting to have to do a lot more work per pixel before they can output a finished product to your monitor, so they'll need a lot more back-end grunt than front-end flash with newer games that touch every pixel with a bunch of shaders, AFAIK.

This is a technical point, though, as I doubt having to touch the majority of pixels in a scene with lights will result in pixels per clock, but rather clocks per pixel render times.

As to the OP, I'm fairly confident that ATi won't catch up to nVidia in Doom 3, simply due to architectural limits. They should get closer (the X800P in particular looks way too slow), but the raw numbers are just stacked in nV's favor.

Well, I went for simplicity and understandability here, but technically you are correct...

 

[Megatomic]

I'll take whatever ATI can give me for my 9800 Pro.

 

[Shamrock]

Originally posted by: GeneralGrievous
"There were no really significant games last year which used Open GL"

This is why there has been no driver rewrite I suppose. And correct me if I am wrong, but aren't the Unreal engines directx?

Really?

Call of Duty was a best seller.

SW: KOTOR was a best seller.

And what is ATI gonna do? step up their "adaptive algorithm" to be even MORE aggressive and lower IQ?

 

[CraigRT]

Originally posted by: reallyscrued
quote : That's bullsh*t

The original Unreal Engine was GLIDE ONLY. Direct3D and OpenGL support was added in patches. Unreal Tournament shifted the engine's rendering focus over to Direct3D, but it was still basically a patch-together hack. Unreal Engine 2 was built from the ground up to take advantage of Direct3D, but it still supports OpenGL rendering.

Unreal Engine 3 will be built with DX9 targeted as the minimum spec, and will support OpenGL rendering.

END QUOTE

whoops, looks like i didnt do my homework. sry man, the unreal i got was coupled with GOTY edition of UT so i got the patched DX version of Unreal and neva knew bout the glide only one back in the day. but so far, i dont think, in all the games based on any unreal engine, gets better performance by picking OpenGL than DirectX. (ive only tested this on ATI cards though.)

Sry quote button dont work....dunno why....ill check it out in forum issues.

I always ran my UT GOTY in openGL, it always seemed to run better than DirectX on my GF2 GTS at the time.

 

[GeneralGrievous]

Really?

Call of Duty was a best seller.

SW: KOTOR was a best seller.

And what is ATI gonna do? step up their "adaptive algorithm" to be even MORE aggressive and lower IQ?

Those are the articles words, not mine. My guess is that ATI thought their CoD performance was fine, and that KOTOR sales were split with the xbox.

I don't know what ATI plans to do, but a speed increase is theoratically possible because they already run directx games much faster.

 

[Keysplayr]

Originally posted by: Insomniak

Originally posted by: keysplayr2003
Ok lemme see if I get this right. Card "A" renders 1/2 as many pixels as card "B" in the same amount of time.

I don't know where the hell you got this from. I never said this, nor did anyone else in this thread that I can see.

Originally posted by: keysplayr2003
Pixels go into frames and the frame can only be rendered when its filled with pixels.

More appropriately, when all the pixels needed to complete the frame have been properly rendered. Different pixel types render different ways on different architectures using different drivers. I could put it differently, but that would be...too different.

Originally posted by: keysplayr2003
Therefore it would take card "A" twice as long to render a frame then card "B".

Here you are back at this two card example again. When did we start talking about cards?

Originally posted by: keysplayr2003
But you say fps increase and pixel render speed are not directly correlated.
You see where I'm confused here right?

Not really. There's a hell of a lot more to rendering a frame than just how fast your pixels come out of the pipe. Familiar with the differences in ARB, ARB2, NV10, NV20 render paths? If not, I'm not going to bother explaining them here. Go google.

Originally posted by: keysplayr2003
it looks like your contradicting yourself in the same sentence.

I don't see how that's the case, considering half the stuff you mentioned up here wasn't said by anyone.





Let me try this again:


Basically, a Doom 3 shadow is an untextured, unlit polygon (for soft, alpha blended onto whatever's behind it). Therefore, in simple terms, the pixel used to render them doesn't have to take up any TMUs when going through a pixel pipeline. It goes pretty much from triangle setup to Z culling to frame buffer.

This is true for both NV4x and R4xx cores. In theory, this means the number of pipelinesx2 = maximum pixel output per clock in Doom 3 PROVIDED THAT:

1) 50% of the pixels being rendered that clock are stencil shadows, and 50% are not.
2) The stencil shadows being rendered are not alpha blended.
3) The non-stencil shadow pixels are able to be rendered in one pass.
4) The drivers/firmware pushing the architectures are equally efficient in their organization and queueing (sp?) of instructions.
5) The fetching of texture data from the memory occurs at equal rates.
6) etc etc etc.


This is why you sometimes hear of NV30 referred to as 8x0 and NV40 referred to as 32x0. 16 pixels per clock with 1 texture per pixel, or 32 pixels per clock with no textures.

Suffice it to say, this situation never happens. But assuming all other things being equal, the number of pipelines a card has directly affects how many pixels it renders per clock in Doom 3.

So my original statement, which is that a card with 16 pipes should outperform a card with 12 pipes, holds true.

The reason performance isn't doubled just by having double the number of pipelines is simple. See 1 - 5 above. Note that the cores of the X800pro and 6800GT (to use two arbitrary examples) are clocked at different speeds. Note that ATi's software suite is obviously not as efficient at driving OpenGL rendering as Nvidia's is. Note 500 other x-factors.

The fact that, all other things being equal, double pipelines = double performance doesn't mean much when in the real world double pipelines does not equal double performance. This is because in the real world, all other things are not equal. However, despite that, having extra pipelines WILL give a given card a significant performance advantage when using the Doom 3 lighting model.

Can you sleep at night now?

Don't get testy my friend. I am just trying to show you that I did not understand what you meant before. Now that I have read your post that you generously written at great length, Its clearer now. I appreciate you taking the time to do it.

I dont know how you picture me when you read my post, but I am not all uptight about this stuff. I am quite calm. Not a ten year old full of angst. Again, I appreciate your explaination.

-Cheers

 

[dfloyd]

"Reviews have showed the X800 Pro losing to the GT big time at high resolution. Or lets say a difference of over 10 fps. Thats huge."

LOL!!!!

Oh how times of changed. I remember the good old days of the Rendition cards and the Voodoo 1. We were super dang impressed to get 20fps. Honestly imo 10fps means nothing. I mean absolutly nothing. Now if the game does not play smoothly that means something. But if your say 100 or 110fps there is no difference other than bragging rights. Now saying 20 or 30 there is a difference, but that goes to the above statement.

And this is not a argument for either card, just laughing at that statement. I crack up so many times that I think people pay so much for 10 more fps. Even though it really does not make any difference in 99% of everything out there. But man those 10fps are worth that extra $200. I love guys like that though honestly. I loved the guy who sold me his 9700 Pro for $150 a year ago so he could go out and buy the $300+ 9800 so he could get 10 more fps.

 

[oldfart]

Originally posted by: dfloyd
Oh how times of changed. I remember the good old days of the Rendition cards and the Voodoo 1. We were super dang impressed to get 20fps.

Yup. I remember tweaking the heck out of my system just to get 20 FPS in Quake2 on my Rendition 2100 card!

 

[BFG10K]

ill believe it when i see it.

That and "why the hell are you optimizing your driver for Doom III when the whole thing needs an overhaul?"

 

[nemesismk2]

With all this talk of ATI promising Doom3 performance boosts reminds me of the the ATI Quack fiasco, if ATI do the same again may I suggest we call it the ATI Dung3 fiasco?

 

[vshah]

Originally posted by: keysplayr2003

Originally posted by: Gamingphreek
Why would it double the performance of a Pro. A Pro has 12 pipes a GT has 16. Where is the doubling there? I certainly see the 16 pipe cards doubling performance of the last gen 8 pipe cards.

Also the 9500 Pro did not have 8 pipes unless you did that wonderful little mod that turned it into a 9700Pro. Do remember that the 5950Ultra was 8 pipes in some situations. When multitexturing for instance it was 8. Also it was clocked much higher than the 9500Pro.

-Kevin

Not for nothing, but can you please read my post again carefully. I said 12 pipe 6800 shoudl have double the performance in DoomIII if the 2 pixel per pipeline is legit using the Z-buffer deal. But it does not and will not have anywhere near double the performance of a X800pro.

the two pixel/pipeline does not work for every pixel in the scene, just a large number. any pixel that still has color info will be rendered as 16x1/12x1

 

[Brian48]

Originally posted by: Insomniak

Originally posted by: reallyscrued
yes unreal engines always have been, and always will be, directX.

is unreal engine 3 9.0C or opted for 10?

That's bullsh*t

The original Unreal Engine was GLIDE ONLY. Direct3D and OpenGL support was added in patches. Unreal Tournament shifted the engine's rendering focus over to Direct3D, but it was still basically a patch-together hack. Unreal Engine 2 was built from the ground up to take advantage of Direct3D, but it still supports OpenGL rendering.

Unreal Engine 3 will be built with DX9 targeted as the minimum spec, and will support OpenGL rendering.

Well, to be absolutely technical about it, the original Unreal engine (aka "Fire" engine as it was called back then) was a software renderer. Glide was added on later in time for the game's launch. Direct3D and OpenGL came later and the engine eventually evolved to the point where they dropped the software component altogether. Although I think DirectX will remain the primary focus for Epic in the years to come, I don't think they will ever completely drop OpenGL so long as they want to maintain some sort of support for the Linux community.