• Will ATI's Tesselation handling get anybetter?
    37 replies, posted
Right the 5000 series seem to be handling tessellation poorly as it seems in the benchmarks. Will a new driver update make it perform better?
Probably not, at least by much. You'd be waiting for the next card refresh for that. IIRC, nvidia has a dedicated engine in their gpu that processes tessellation. Might be wrong on that.
It's not like games will use such a insane amount of tessellation.
[QUOTE=Brage Nyman;22128599]It's not like games will use such a insane amount of tessellation.[/QUOTE]This. The amount of Tessellation in that Heaven demo was obscene.
Why don't the game makers just make games look better instead of relying on Gpu manufacturers?
[QUOTE=Blarg190;22128745]Why don't the game makers just make games look better instead of relying on Gpu manufacturers?[/QUOTE] Less people would be able to run it, which means less sales?
[QUOTE=Blarg190;22128745]Why don't the game makers just make games look better instead of relying on Gpu manufacturers?[/QUOTE] Because you can't make it better if only one or two people can run it.
i mean come on tessellation is pretty damn hyped, heavens demo can't even make a couple of stairs with tessellation instead they make one huge ramp with textures on it. Even half-life was able to make fucking stairs.
[QUOTE=Brage Nyman;22129049]i mean come on tessellation is pretty damn hyped, heavens demo can't even make a couple of stairs with tessellation instead they make one huge ramp with textures on it. Even half-life was able to make fucking stairs.[/QUOTE] You don't know what tessellation is for, do you?
[QUOTE=JWJ;22129124]You don't know what tessellation is for, do you?[/QUOTE] Talking out of his ass.
[QUOTE=Blarg190;22128745]Why don't the game makers just make games look better instead of relying on Gpu manufacturers?[/QUOTE] Easier said than done. You can't always achieve the detail you want because of time restrictions. When you make models for characters and whatnot, you also need need to make 1 to 3 LOD models for that character or object that have less detail. That is so extra detail that is not necessary won't be rendered from far away. Additionally, depending on whether the game is multi-platformed, you might have to adjust the models so they are catered to their respective system. Tesselation removes this mess. Developers only need to make one model and the GPU can easily control the level of detail independently. This cuts down on time in development a lot. Also, you can notice in games sometimes when objects change to their lower detail LOD model counterparts because they "pop" or, better put, you can see finer details on the object simply disappear. GPU tesselation is more subtle and you will commonly never notice the changes in models. So, with GPU tesselation, you get: - Less time spent optimizing your game. - Less visual quirks. I also agree with most users. The "Heaven" demo's tesselation is way over-the-top. I've seen it in wire-frame and it's insane. Given that, I don't think it should manipulate your decision on which kind of video card to buy. But know that Nvidia does indeed do tesselation better due to it's built in dedicated tesselation hardware. In the future, tesselation will surely become as important as what pixel shaders are today.
[QUOTE=JWJ;22129124]You don't know what tessellation is for, do you?[/QUOTE] You don't know what heavens benchmark is? Been living under a rock? ^^ [img]http://pici.se/pictures/GRbYOsinM.jpg[/img] Look at those stairs! They really look like shit without tessellation! Oh my we really need tessellation to make stairs!
[QUOTE=that1dude24;22128393]Probably not, at least by much. You'd be waiting for the next card refresh for that. IIRC, nvidia has a dedicated engine in their gpu that processes tessellation. Might be wrong on that.[/QUOTE] ati uses a dedicated tessellation engine. nvidia uses the cuda cores for tessellation. And yes, apparently they will increase tessellation for southern islands.
[QUOTE=Brage Nyman;22129577]You don't know what heavens benchmark is? Been living under a rock? ^^ [img_thumb]http://pici.se/pictures/GRbYOsinM.jpg[/img_thumb] Look at those stairs! They really look like shit without tessellation! Oh my we really need tessellation to make stairs![/QUOTE] That is what is great about tesselation, that is all the developer had to make, a simple ramp looking piece of shit. And with a displacement map texture, the tesselation engine takes care of the rest. That is much easier than having a modeler make a detailed set of stairs, and then making one or two additional LOD models and UV wrapping them individually.
He is being complicated.. we don't need to argue any further.. I'm sure ATI will improve tessellation performance, but there is only so much you can do
No, it won't be improved, not significantly anyway. It's on the hardware level.
[QUOTE=Brage Nyman;22129577]blah[/quote] Do you realize it's a showcase, thus they make shitty looking stairs to show how awesome they look with tesselation?
Considering Cypress was the first DX11 card to be released back in September, they really didn't make it specifically to process heavy tessellation to begin with. It is hardware based and has a [i]fixed[/i] performance on all the high, medium, and low end series cards (for the most part) regardless of how many shaders are onboard. Of course, clock speed and memory bandwidth still apply. Tests revealed that its tessellation performance was almost resolution independent, which ruled out a suspicion of it being the rendering pixels that clogged everything up. Read more here for a better understanding, it explains it more in depth. [url]http://www.beyond3d.com/content/reviews/54/9[/url] Fermi was basically taken to a whole new level in this sense; the tessellation power is a function of how many shader cores are in the GPU. Performance will scale downward linearly as the shader core count decreases in mainstream and budget cards. Of course, the lower GPU clock speed will decrease it slightly more, but not by much. Unfortunately, it wasn't able to reflect this monstrous performance advantage as seen in synthetic benchmarks like Uniheaven because in real world situations the shaders used for tessellation will be doing other things needed to render pixels on the screen than just primarily for processing tessellation. Uniheaven benchmark is a perfect example of this; it is almost a pure tessellation bench more than anything else out there currently. No other bench or game can match the power needed in this benchmark. After I figured out how Fermi's tessellation worked I wasn't surprised to see it perform so well in Uniheaven. Unfortunately I went about explaining it in the completely wrong way like I do with most things for the first time and usually end up fucking up badly and causing a brawl in the process most of the time getting banned for various reasons (Very bad habit in me...feel free to have a good laugh looking at some of them)....anyways that's a whole different subject that we can get into another time To simplify, Fermi has a polymorph engine in each Streaming Multiprocessor block that works in conjunction with the associated SP array; basically this distributed form of tessellation gives it enourmous peak tessellation performance. The downside to this is that it uses the shader array to do it. Not that it's a bad thing of course, it just means the performance won't exactly be....consistent...depending on the game/bench/rendering method/video settings/etc umm..well....getting onto the question about ATI...the Cypress refresh is obviously going to have to have some significant increases in tessellation performance as I'm sure AMD is quite well aware of what's happening in that department with their competition currently. I don't know how, they might even redo the whole thing from scratch and use a different method. Implementing methods of tessellation into hardware is quite easy as pretty much every part in a GPU can do it; the question is what is the most effective method. The Cypress ASIC has lots of room for improvement so I could easily see something positive in the refresh, but I personally don't use tessellation at all on my 5870 (COP/Metro) because I hardly notice a visual difference and quite frankly, could care less about until maybe 1 or 2 years down the road. To sum it all up, Fermi's design is more flexible and has a higher peak performance, while Cypress's design is more consistent and will scale better with the lower-end families due to the tessellator being a fixed function.
I have a question and it sounds dumb and forgive me for not knowing that much about tesselation: If the customers need GPU's that support tesselation to play a game, wouldn't that result in less sales? If so, wouldn't it just be better to make those stairs? [editline]10:53AM[/editline] I don't really know anything about tesselation so answer my question by explaining how tesselation works really. [editline]10:54AM[/editline] Well I know how tesselation works in the MC escher way but not how it's incorporated with hardware.
[QUOTE=Hostel;22130118]That is what is great about tesselation, that is all the developer had to make, a simple ramp looking piece of shit. And with a displacement map texture, the tesselation engine takes care of the rest. That is much easier than having a modeler make a detailed set of stairs, and then making one or two additional LOD models and UV wrapping them individually.[/QUOTE] Because it is so hard to make a rectangle and place another one above and behind it and loop that. And stairs wouldn't have to be made into a model, it can be done with world geometry.
Outside of heaven, I havent really seen large use of tesselation.
[QUOTE=Dark-Energy;22139080]Considering Cypress was the first DX11 card to be released back in September, they really didn't make it specifically to process heavy tessellation to begin with. It is hardware based and has a [i]fixed[/i] performance on all the high, medium, and low end series cards (for the most part) regardless of how many shaders are onboard. Of course, clock speed and memory bandwidth still apply. Tests revealed that its tessellation performance was almost resolution independent, which ruled out a suspicion of it being the rendering pixels that clogged everything up. Read more here for a better understanding, it explains it more in depth. [url]http://www.beyond3d.com/content/reviews/54/9[/url] Fermi was basically taken to a whole new level in this sense; the tessellation power is a function of how many shader cores are in the GPU. Performance will scale downward linearly as the shader core count decreases in mainstream and budget cards. Of course, the lower GPU clock speed will decrease it slightly more, but not by much. Unfortunately, it wasn't able to reflect this monstrous performance advantage as seen in synthetic benchmarks like Uniheaven because in real world situations the shaders used for tessellation will be doing other things needed to render pixels on the screen than just primarily for processing tessellation. Uniheaven benchmark is a perfect example of this; it is almost a pure tessellation bench more than anything else out there currently. No other bench or game can match the power needed in this benchmark. After I figured out how Fermi's tessellation worked I wasn't surprised to see it perform so well in Uniheaven. Unfortunately I went about explaining it in the completely wrong way like I do with most things for the first time and usually end up fucking up badly and causing a brawl in the process most of the time getting banned for various reasons (Very bad habit in me...feel free to have a good laugh looking at some of them)....anyways that's a whole different subject that we can get into another time To simplify, Fermi has a polymorph engine in each Streaming Multiprocessor block that works in conjunction with the associated SP array; basically this distributed form of tessellation gives it enourmous peak tessellation performance. The downside to this is that it uses the shader array to do it. Not that it's a bad thing of course, it just means the performance won't exactly be....consistent...depending on the game/bench/rendering method/video settings/etc umm..well....getting onto the question about ATI...the Cypress refresh is obviously going to have to have some significant increases in tessellation performance as I'm sure AMD is quite well aware of what's happening in that department with their competition currently. I don't know how, they might even redo the whole thing from scratch and use a different method. Implementing methods of tessellation into hardware is quite easy as pretty much every part in a GPU can do it; the question is what is the most effective method. The Cypress ASIC has lots of room for improvement so I could easily see something positive in the refresh, but I personally don't use tessellation at all on my 5870 (COP/Metro) because I hardly notice a visual difference and quite frankly, could care less about until maybe 1 or 2 years down the road. To sum it all up, Fermi's design is more flexible and has a higher peak performance, while Cypress's design is more consistent and will scale better with the lower-end families due to the tessellator being a fixed function.[/QUOTE] China called, they want their great wall back.
Next refresh of cards will probably have better tesselation. But I don't think we'll see any games using even what the 5000 series can do before the next generation/refresh is out from both nvidia and ati.
[QUOTE=G-Strogg;22140894]I have a question and it sounds dumb and forgive me for not knowing that much about tesselation: If the customers need GPU's that support tesselation to play a game, wouldn't that result in less sales? If so, wouldn't it just be better to make those stairs? [/QUOTE] No, the benchmark shows what tessellation can do, not what it should be used for. It's up to developer to make stairs or to use displacement and tessalation to create them. [QUOTE=Karmah;22145503] And stairs wouldn't have to be made into a model, it can be done with world geometry.[/QUOTE] Most engines have poor/no world geometry support. So no. Anyway, tessalation will NOT replace LOD. While lod reduces quality, tessalation can only add triangles, not remove them.
[QUOTE=johanz;22148822] Anyway, tessalation will NOT replace LOD. While lod reduces quality, tessalation can only add triangles, not remove them.[/QUOTE] What? As you get closer, tessellation adds more triangles, so obviously if you back away it [i]will[/i] remove them.
[QUOTE=Hostel;22130118]That is what is great about tesselation, that is all the developer had to make, a simple ramp looking piece of shit. And with a displacement map texture, the tesselation engine takes care of the rest. That is much easier than having a modeler make a detailed set of stairs, and then making one or two additional LOD models and UV wrapping them individually.[/QUOTE] Uh no. It takes longer (or just as long if you really want your stairs to look good) to tessalate than it does to just model fucking stairs. Wonder how they get tessalateable (or any) normal maps for objects? They model them. 99% of bump maps, no matter how basic (standard bump maps) to advanced (tessalation, which is the same thing except actually projects it out into the world) come from actually modeling out the detail. That's why artists make high poly models of things even if a game can only handle 1/5th of the detail in said object. They make a really high poly model, export all that detail in the model into a normal map or bump map, and then wrap it around a low poly object so it gives the illusion that it is high poly. Tessalation works EXACTLY the same art-wise... except the rendering engine obviously does a much more detailed job with it so it looks better. Those detailed bump maps have to come from somewhere, and usually it comes from it being modeled out. At least for objects. [editline]08:57PM[/editline] The only thing tessalation does that might save time is that you don't need to make seperate lod objects for each model you make in order to save performance, assuming everyone who runs your game has a Dx11 card.
[QUOTE=Brage Nyman;22129577]You don't know what heavens benchmark is? Been living under a rock? ^^ [img_thumb]http://pici.se/pictures/GRbYOsinM.jpg[/img_thumb] Look at those stairs! They really look like shit without tessellation! Oh my we really need tessellation to make stairs![/QUOTE] Maybe it suppose to be a ramp? :downs:
the problem w/ the haven demo is they didn't make any effort to make the image quality competitive when not using tessellation. its not a fair and generous comparison so its void. in a real game w/o tessellation, the artists will use some crazy hacks to make it look nice and also run on dx9 hardware. its the same problem w/ the 'unlimited detail' videos.
[QUOTE=B1N4RY!;22151240]Maybe it suppose to be a ramp? :downs:[/QUOTE] or maybe they could had made like 10 somewhat square steps in dx 9 / 10 and don't overhype it. At least the ramps is good for handicapped people like me.
[QUOTE=Odellus;22148921]What? As you get closer, tessellation adds more triangles, so obviously if you back away it [i]will[/i] remove them.[/QUOTE] If you create a character with 10k tris, tessalation will NOT make it 8k if you back away.
Sorry, you need to Log In to post a reply to this thread.