Half-Life 2 XB360 won't be using tiling to achieve A.A.

[Bobbler]

What an absurd rebutal!

I called it absurd because the implication that developers will be programming to RSX's weakness for some reason, vs. it's strengths (therefore only showing RSX's weaknesses compared to Xenos). That is, frankly, absurd. I'm sure we'll see lots of amazing stuff from both systems, but to say one system will show strenghts while the other will show weakness is sort of a wonky view of things.

 

[kimg]

Just looking at ATI's X1000 series and Xenos and where D3D10 is headed, and how RSX was implimented into the PS3, I can see a number of areas where RSX is going to show limitations in some future 3D game engines. RSX may be an SM3.0 GPU, but is has very poor dynamic branching in pixel shaders and its vertex texturing is very slow. IMO, It is an SM2.0 GPU with SM3.0 features stapled on. It really doesn't do any core SM3.0 feature exceedingly well. Further, RSX has a split memory pool, is facing texturing penalties from XDR, has half the fillrate a typical GPU with its texturing and shading power possesses, has a significant performance penalty for FP16 blending and filtering, and while it has a paper flops advantage it has less effecient ALU utilization and also ties up shader ALUs when texturing.

Not that Xenos is perfect--not by a long shot. Tiling has been an issue (too small for 720p 4xMSAA, requires a renderer rewrite if there is not an early Z pass, predicated tiling support was not really resolved until Spring 2006, memexport requires a 3x load on the CPU and was not compatible with tiling, etc) and MS did not resolve it very quickly, MS did not get SM3.0 capable GPUs into developer hands until nearly the consoles launch, the dev kit situation was fubared and illconceived, Xenos is looking 2 years down the road and has a significant amount of transistor budget allocated in a way that doesn't help a lot of current game engines, has less peak texel fillrate, etc.

There are hurdles on both sides. But I agree with Megadrive: When developers begin developing SM3.0-based engines and begin exploring basic D3D10 features there is going to be decided overlap with the core focus of Xenos.

Of course I can see the arguement now: But the PS3 can offload work to Cell. Which is true (of both consoles), but the topic is the GPUs, not the systems.

Im sorry if i disapoints you, but xenos do not fulfill the demands of a dirextx10 gpu to my knowledge. It is shader model 4.0 who is demanded by DX10, what will probably get ported(hope you knew this, and were just messing up your post), is probably the dirextx9 part of Vista's graphics model, if applicable.

A lot of graphics cards can exploit the power of the new Directx implemented in Vista, including my own geforce 6800xt. Im aknowledging xenos got a good architecture, but only time will tell how it REALLY compare to a true directx10 gpu with SM4.0

 

[AlNom]

Im sorry if i disapoints you, but xenos do not fulfill the demands of a dirextx10 gpu to my knowledge. It is shader model 4.0 who is demanded by DX10, what will probably get ported(hope you knew this, and were just messing up your post), is probably the dirextx9 part of Vista's graphics model, if applicable.

He never says that Xenos is a D3D10 part. Xenos' features are more than SM3.0, but it also lacks certain features of D3D10, hence mentioning "basic D3D10 features".

 

[Shifty Geezer]

thats why i believe they will use memexport, you cant use tiling when using memexport and you cant use memexport when using Tiling so I believe when taking into the concept and features of using memexport, Valve could use this implementation to solve the AA problem without using AA.

You know, you could potentially blow this board away if you can explain your method. The ROPs are in the eDRAM and use the phenominal BW of the eDRAM to enable MSAA wihtout eating into the rest of the system's BW. If you're not using the eDRAM for AA, you're not using the ROPs and you're working from main RAM. MEMEXPORT allows you to arbitarily write out data from within a shader to any memory location.

So, what s this amazing method you have to write values to RAM such that you can add AA, without getting tied up with the other bandwidth demands of the system all sharing that 22 GB/s bus? Please, explain the process in detail, including approximate bandwidth consumption.

 

[Farid]

screenspace vector based effect? would you be so kind as to explain or point me in a direction to find info on this?

I can do both. I'm cool just like that.
Well, I can explain it simply, given that the papers and presentations linked in this post go into the gritty details already.

MB is obtained in Real-Time by rendering the frame you want to see motion blurred first and then an offset buffer, or velocity buffer, in second. Offset buffer which is calculated on a per vertex basis (according to the movement direction) and then stored in screenspace (RGB, like a Normal Map, if you will). And finally you apply a blur on the pixels of your original frame based on the vectorial (movement) information stored in the offset buffer.

And as they say a picture worth a thousand words, so here's some pictures to illustrate that:

You render the frame you want to se motion blurred, in this case a tunnel:

You calculate the velocity information and then buffer it, store it into screenspace RGB information, this gives you something like this (without the pointing arrows, of course ) :

And finally you apply your blur effect (akin to the ones you have in Photoshop, Matrix Convultion and the like) to the frame. the amount of blur applied per pixel, relatively to the other pixels of the frame, is determined by your velocity buffer:

Pictures shamelessly stolen from an Ati presentation and uploaded to xs.to.

I hope that this explanation was of any use to you, or to anyone else.

And here's the promised paper and presentations:
An academic paper on a RT implementation.
http://www.dtc.umn.edu/~ashesh/our_papers/motion-blur.pdf
A sample article from the excellent ShaderX book series, from an implementation from Ati guys:
http://www.ati.com/developer/ShaderX2_MotionBlurUsingGeometryAndShadingDistortion.pdf
An Ati presentation that sum up the technique in more simplistic terms:
http://www.ati.com/developer/gdc/GDC2005_BringingHollywoodToRealtime.pdf

You have other sources on the subject in the references section of these papers.

Valve, who are clearly more superior developers than you guys, no pun intended, can achieve the same effect with AA on 360.

How charming and baseless at the same time.
May I suggest you not to be rude with folks just because you can't address the correct technical points they make?

i believe when you have unified architecture with unused features like memexport and an API which is inbetween dx9 and dx10 still in its infancy compared to an already established GPU related API of Nvidia, it would be easier for devs to show with Nvidia thier strengths eaarly and the ones with ATI Xenos, theier strengths later

That has stricly nothing to do with operating Full Screen Anti Aliasing - on a 720p RenderTarget - comparable in quality to SSAA and MSAA on C1/Xenos without having recourse to tiling. Nothing.
Actually, it barely makes any sense.

i believe you are wrong in this one

Where is your technical explanation behind this belief of yours. Is that just blind faith that push you to believe Valve could obtain an AA as good as RGMSAA while having a smaller performance hit than simply tiling the framebuffer into 3 parts...

just like you guys developed a system of not wasting system resources and got a form of AA with faked hdr at the same time keeping the visuals intact

Firstly, this form of AA Heavenly Sword uses is hardware supported and called MSAA, secondly what is your definition of High Dynamic Range Imaging and Rendering, exactly?

Because as far as I know NAO32 has the capability to store image information superior and inferior to [0,1.0], just like other HDR implementation do.

You can implement HDR in Real Time by different manners, using FP16 RGBa RenderTarget is just one of them. FP10 RT and NAO32 (using a different Colour Space) are just different implementations, not "fake" and/or incorrect ones. Or else you could just as well consider any format but one using 4096bits per pixels, or something, as being a "fake" one.

There's only difference in maximum precision and implementation limitations between the various HDRR available. Nothing to do with true or fake.

thats why i believe they will use memexport, you cant use tiling when using memexport and you cant use memexport when using Tiling so I believe when taking into the concept and features of using memexport, Valve could use this implementation to solve the AA problem without using AA.

So, simply put, you like the term MEMEXPORT as a buzzword and you try to shoehorn it into the discussion? No matter how little relenvancy it has.

 

[Titanio]

You render the frame you want to se motion blurred, in this case a tunnel:

You calculate the velocity information and then buffer it, store it into screenspace RGB information, this gives you something like this (without the pointing arrows, of course ) :

And finally you apply your blur effect (akin to the ones you have in Photoshop, Matrix Convultion and the like) to the frame. the amount of blur applied per pixel, relatively to the other pixels of the frame, is determined by your velocity buffer:

Pictures shamelessly stolen from an Ati presentation and uploaded to xs.to.

I hope that this explanation was of any use to you, or to anyone else.

This was good, thank you I've one question - how do you calculate the velocity vectors for each vertex, for static geometry? Do you actually move the scene around the camera (and keep the camera static), and just use that velocity directly, or do you use the camera's velocity and calculate a velocity vectors for the scene's vertices based on that, and the mirror of the camera's vector's direction (or something like that..?)?

edit - I guess you'd calculate the velocity vectors based on both camera and object movement. If you had one object going in a direction at a certain speed, you'd have a vector in that direction. If the camera was going at the same speed in the same direction, you'd have the mirror. Adding, you'd have a net velocity vector of 0,0,0, which is what should be the result (no blurring).

 

[Jesus2006]

Just looking at ATI's X1000 series and Xenos and where D3D10 is headed, and how RSX was implimented into the PS3, I can see a number of areas where RSX is going to show limitations in some future 3D game engines. RSX may be an SM3.0 GPU, but is has very poor dynamic branching in pixel shaders and its vertex texturing is very slow. IMO, It is an SM2.0 GPU with SM3.0 features stapled on. It really doesn't do any core SM3.0 feature exceedingly well.

Do you have any technical or experienced evidence for these claims?

 

[Acert93]

I called it absurd because the implication that developers will be programming to RSX's weakness for some reason, vs. it's strengths (therefore only showing RSX's weaknesses compared to Xenos). That is, frankly, absurd. I'm sure we'll see lots of amazing stuff from both systems, but to say one system will show strenghts while the other will show weakness is sort of a wonky view of things.

By this reasoning DX9 or D3D10 were never necessary because we could have continued exploiting DX8 and just emphasized what was already available. But this ignores the benefit of featureset and how it impacts final quality. You can exploit DX8 as much as you want, but DX9 features when equally exploited will look better.

Don't believe me? A perfect example of similarly performing GPUs, FX5900 and Radeon 9800, showed this with DX9. The DX5900 was a good DX8.1 performer and easily on par with the Radeon, but when you used a DX9 SM2.0 path the Radeon was a good 2x faster. These GPUs had similar raw performance, but when it came to feature and feature performance in SM2.0 the FX was out of its league.

And this is the case with Xenos and RSX in many regards. Xenos is a much, much better GPU in regards to dynamic branching in pixel shaders and has quite a few functional features that RSX either lacks or is very slow at for the very reason it SM3.0 features are not the core target of the processor.

Obviously it is easy to say this arguement is absurd since SM3.0, tesselation, and other effects are not in the mainstream baseline yet. The operative word is YET. You need look no further than the Fall of 2002 when review sites took your VERY arguement. And when Spring 2004 hit (Far Cry), and later Fall (HL2), the disparity in visuals was significant for the very reason that the functional performance of the Radeon in "features soon, but not yet utilized" was better.

Of course PS3 developers will avoid these areas where necessary, as you note. The question is whether you believe, on a GPU level, whether being fast at SM3.0 features can give added performance and/or improved IQ over a slower SM3.0 GPU. I believe history says, on the GPU level, that the better featureset will show fewer limitations down the road regardless if you try to avoid the pit falls for the very reason the additional featureset gives improved image results.

Im sorry if i disapoints you, but xenos do not fulfill the demands of a dirextx10 gpu to my knowledge.

I never said as much. What I said:

When developers begin developing SM3.0-based engines and begin exploring basic D3D10 features there is going to be decided overlap with the core focus of Xenos.

And yes, Xenos does have some overlap and continuity with some of the core focus of D3D10. Of course it lacks GS in name, lacks interger support in the shaders, etc but there is a reason MS requested the design from ATI and that ATI is leveraging the basic concept of the part for their D3D10 offering.

Do you have any technical or experienced evidence for these claims?

Check out the PC sections of this site. This is pretty standard for NV in regards to adding new architecture check boxes and the feature support and performance of the NV40/G70 line in core SM3.0 features like vertex texturing, dynamic branching in pixel shaders, etc is pretty poor. e.g. There are times when an X1600 can outperform a GF7900 when dynamic branching in Pixel Shaders are used (e.g. POM). In a number of shader benchmarks using dynamic branches the X1900 comes out up to 10x as fast as the 7900. This is due to the finer granularity of the batches (e.g. X1900 batch size is 48; G70 is >1000; Xenos batch size is 64). Vertex texturing is another example; G70 has significant latency while texturing from vertex shaders and only has 8 of them to begin with. Xenos obviously can dynamically load balance (up to 48 vertex shaders), but since the ALUs are "generic" all 48 can texture -- and with the same 'penalty' as texturing from a pixel shader (which is small). This is one benefit of having the texturing units decoupled from the shader pipeline (e.g. in RSX you use up some shader resources while texturing). Xenos also has 16 dedicated point sample units in the TMUs with the core focus of vertex texturing.

If such features bring better image quality, which I believe they can, Xenos has an edge in these areas in addition to things like fillrate, certain aspects of the framebuffer/bandwidth, and features like tesselation and FP10 (which is decidely faster than FP16). Not that Xenos is perfect (as I noted above) but there are forward-looking features (both DX9 SM3.0 and D3D10) included in Xenos that perform well and as time goes on game engines will begin relying more and more on these techniques.

 

[scificube]

removed.

 

[Farid]

Keep on Topic, folks.

Why does this thread turn into yet another X360 Versus PS3 thread, exactly?

 

[scificube]

Why does this thread turn into yet another X360 Versus PS3 thread, exactly?

It shouldn't have. It's not X360 vs. PS3 for me, but I'll remove my comments in any case.

 

[Dr. Nick]

At first I was very excited about all the features of the Xenos but then I realized it might be years before we see the benefits of it or we might never see some of these feature used to their potential in games.
On the Xbox 1 you had DOA3 which made good use of some of the GPUs features but didn't really make use of lighting and then you have Splinter Cell which only really had good lighting everything else was graphically blah. It wasn't until late in the systems life when devs started to make good use of its features and that was probably because many of these features were being used in more recent PC games. With the Xenos this might be the case and a feature like MEMEXPORT might never be used because as far as I know it will never be supported in PC games.

 

[Bobbler]

By this reasoning DX9 or D3D10 were never necessary because we could have continued exploiting DX8 and just emphasized what was already available. But this ignores the benefit of featureset and how it impacts final quality. You can exploit DX8 as much as you want, but DX9 features when equally exploited will look better.

Don't believe me? A perfect example of similarly performing GPUs, FX5900 and Radeon 9800, showed this with DX9. The DX5900 was a good DX8.1 performer and easily on par with the Radeon, but when you used a DX9 SM2.0 path the Radeon was a good 2x faster. These GPUs had similar raw performance, but when it came to feature and feature performance in SM2.0 the FX was out of its league.

And this is the case with Xenos and RSX in many regards. Xenos is a much, much better GPU in regards to dynamic branching in pixel shaders and has quite a few functional features that RSX either lacks or is very slow at for the very reason it SM3.0 features are not the core target of the processor.

Obviously it is easy to say this arguement is absurd since SM3.0, tesselation, and other effects are not in the mainstream baseline yet. The operative word is YET. You need look no further than the Fall of 2002 when review sites took your VERY arguement. And when Spring 2004 hit (Far Cry), and later Fall (HL2), the disparity in visuals was significant for the very reason that the functional performance of the Radeon in "features soon, but not yet utilized" was better.

Of course PS3 developers will avoid these areas where necessary, as you note. The question is whether you believe, on a GPU level, whether being fast at SM3.0 features can give added performance and/or improved IQ over a slower SM3.0 GPU. I believe history says, on the GPU level, that the better featureset will show fewer limitations down the road regardless if you try to avoid the pit falls for the very reason the additional featureset gives improved image results.

My point really had nothing to do with Xenos' specs -- it was about the implication that only weaknesses will show up with RSX. It is absurd, and regardless of how many pages you write repeating it, it doesn't mean Xenos' dynamic branching and such make it so RSX can't show render stuff without showing weaknesses compared to Xenos. Xenos does not have RSX beat everywhere, which seems to be what you're thinking. It's very much arguable that we've yet to see any game take full advantage of RSX style GPU even -- you can say it's nice that RSX is straight from a PC design so devs are used to it, but that's not quite true... there isn't a any PC devs out there that program straight to the GPU, knowing it's faults and strengths like in a closed box design (PC devs are essentially programming to the API, not the hardware).

It's akin to saying that because Xbox is clearly superior to PS2 in features, PS2 will only show weaknesses. You don't see the oddity in what you're defending? Xenos may be good at a lot of new fancy features, but it doesn't mean RSX is all of a sudden a dud that is only capable of running in second place to xenos -- maybe that's true, but I've seen nothing to support such claims, regardless of dynamic branching. It isn't a case where Xenos is pantently better at everything and so we can claim that it's few features that its excellent at will propel it into a state where nothing RSX can do could make a picture as pretty (in a different way). You say that Xenos' features will show some new exciting things that RSX can do, this may be true, but does that mean RSX can't do anything better than Xenos? I don't know to be honest, but it seems strange to me to argue that one will show itself to be patently better than the other over time, when that has never happened in any console generation, ever -- there have been things that one was better at than the other, sure, but not like you or Megadrive's initial comment seemed to be saying. Maybe I'm arguing symantics, but it still seems silly to me.

 

[Sojurn]

I don't think their comment about efficiency was specifcially referring to the Xbox 360, But traditional GPU. The short of it is, They are not willing to spend money doing things for 360 which would be useless for PS3 and PC. If it means the 360 version is the worst, so be it.

 

[Fafalada]

Ok, I will bite

Try to bite in context then.
We've been told for last 2 years how Xenos is the staplemark of efficiency (especially in shader usage) while RSX requires special case attention/work to get utilization up, which is a direct contradiction to what Megadrive inferred in his post, hence nAo's sarcastic reply.

I believe history says, on the GPU level, that the better featureset will show fewer limitations down the road

I've seen enough cases that say different, even in very recent history.
I know lots of people like to play checkbox comparisons, but in the realworld accumulating longer feature lists doesn't necesserily work in favour of end results.
Unfortunately(for consumers and developers alike) marketting departments have yet to learn that...

 

[Jesus2006]

Try to bite in context then.
We've been told for last 2 years how Xenos is the staplemark of efficiency (especially in shader usage) while RSX requires special case attention/work to get utilization up, which is a direct contradiction to what Megadrive inferred in his post, hence nAo's sarcastic reply.

That's only true. However, i still wonder how developers seem to have so little problems with pushing out up to 4xAA (even at 60 fps, Warhawk) on PS3, while bandwidth is only half of the current high end graphic cards.

On the contrary, the 360 hardly does this, although we are already in 2nd gen now. Instead some games even work at sub 720p resolutions to avoid tiling and get a considerable performance.

Imo this is the best indication for the capabilities of the hardware. The Xenos has alot of paper spec plusses, but it's seriously limited by other factors as it seems.

 

[Fox5]

I sent a email to Valve asking about A.A. and tiling and got a response.



So it will have A.A. but it won't be using the tiling method.

I'd imagine it will use it the same way the PC version did, selective application to surfaces deemed troublesome.

 

[DuckThor Evil]

Hmm I just wanted to say that: Hey! I like 360 or actually I like playing games on it downloading stuff from Live etc., but how much longer do we have to wait before it can be
said that the desing choice to use 10MB Edram has turned out to be nearly a catastrophe!

So because eDram is only 10MB you need to tile the picture if you wan't that advertized 4xAA that was originally supposed to be in every game. I asked few months ago that are we ever going to see AA and 720p in multiplatform games on X360 and I was concerned that it would be a big deal, however nobody answered anything to that. Now few months later it seems that many exlusives won't tile either, atleast the ones that use UE3, and to be fair that engine is used a lot!

I'm not very technical person and I'm sorry if I have missed some vital points about this and I would certainly appreciate if someone could point to me what I'm missing, but at the moment it really seems to me that MS's and ATI's desing choice has turned out to be less that what it was supposed to be. I mean even if there are going to be first party games that do things the way it was meant to be done(tiling to get 720p with 4xAA and stuff!) it's hardly worth it, if at the end of the day there are 5 exlusives that actually use the method and 25 exlusives that don't on top of all the multiplatform titles.

I know the "Sun" hasn't gone dawn yet on the matter, but it's been up quite some time now and things are starting to bother me.

 

[Hardknock]

That's only true. However, i still wonder how developers seem to have so little problems with pushing out up to 4xAA (even at 60 fps, Warhawk) on PS3, while bandwidth is only half of the current high end graphic cards.

On the contrary, the 360 hardly does this, although we are already in 2nd gen now. Instead some games even work at sub 720p resolutions to avoid tiling and get a considerable performance.

Imo this is the best indication for the capabilities of the hardware. The Xenos has alot of paper spec plusses, but it's seriously limited by other factors as it seems.

"have so little problems"? You've got to be kidding me? You're acting like Warhawk is some new graphical benchmark. I would expect a game that looks like Warhawk to be 4xAA and 60fps. RR6 is 4xAA + 60fps and DOA4 is 1080i 4xAA and 60fps.

Developers have repeatedly stated the systems are within arms reach of each other. Stop making this a bigger issue than it is. Compare Resistance (30fps) to COD2 (60fps) does that automatically make 360 twice as powerful as PS3? Of course not!

 

[hey69]

RR6 is 4xAA + 60fps and DOA4 is 1080i 4xAA and 60fps

not that these games are pushing the GFX department ...