City Of The Future Tessellation demo

[Scali]

Question. Is it possible Tesselation could allow Fusion to perform better with limited memory bandwidth?

Depends on how good your tessellation unit is.

 

[cusideabelincoln]

Ultimately, the only way to present realism in a game is to incorporate physics characteristics in games so that objects can react as they do in the real world.

Let's hope AMD and NV both continue to push tessellation and physics (not PhysX) in the future because these demos do look very promising.

I agree. I want to see interactive physics. I like the direction of Portal 2 and its Fluid Dynamics.

 

[flopper]

Exactly, which is why I wanted to show a 'proper' example of tessellation.
For some games it's easier to add 'proper' tessellation than others. Games which use generated geometry anyway, such as for large terrains, like Civ5 and HAWX 2.

As you can see, you really need to design your geometry with tessellation in mind. The geometry needs to be low-detail, but with enough information to scale it up to extremely detailed close-ups using dynamic tessellation.



No, but I think they won't want to go below the level of Crysis/Warhead with Crysis 2. Their engine is capable of scaling very well anyway.

what difference in fps can one expect using such approach where the geometry is design with tess in mind?

 

[Scali]

what difference in fps can one expect using such approach where the geometry is design with tess in mind?

Why are people thinking in terms of fps and such? You really need to stop doing that.
This example shows 1.3 billion triangles on screen. You can't even fit that amount in memory in the first place. Your average vertex takes about 32 bytes in memory, at the very least (position, normal vector, texture coordinates, that's 3 + 3 + 2 = 8 float values, 8*4 bytes = 32 bytes).
A triangle takes 3 vertices.
So we would need 1.3b * 3 * 32 = 124 GB of memory to store this geometry.

As they say, this is about 500x(!) more than what games have so far (Crysis averages around 2M polys on screen at the highest detail settings).

Basically this is simply impossible to do without tessellation. What's the framerate for impossible?
And THAT is what AMD is worried about. Tessellation is the only way forward for more detail in games. And AMD cannot do this. They want developers to lower the tessellation factors, and use more video memory.
With a proper tessellation implementation in hardware, we can do this NOW, as nVidia demonstrates. This is what DX11 is supposed to do.

 

[flopper]

Why are people thinking in terms of fps and such? You really need to stop doing that.
This example shows 1.3 billion triangles on screen. You can't even fit that amount in memory in the first place. Your average vertex takes about 32 bytes in memory, at the very least (position, normal vector, texture coordinates, that's 3 + 3 + 2 = 8 float values, 8*4 bytes = 32 bytes).
A triangle takes 3 vertices.
So we would need 1.3b * 3 * 32 = 124 GB of memory to store this geometry.

As they say, this is about 500x(!) more than what games have so far (Crysis averages around 2M polys on screen at the highest detail settings).

Basically this is simply impossible to do without tessellation. What's the framerate for impossible?
And THAT is what AMD is worried about. Tessellation is the only way forward for more detail in games. And AMD cannot do this. They want developers to lower the tessellation factors, and use more video memory.
With a proper tessellation implementation in hardware, we can do this NOW, as nVidia demonstrates. This is what DX11 is supposed to do.

Ok I get that, fps is what counts especially playing shooters which you might not do.

However, if I understand you right, we can add more detail and create a lush for example jungle, which is even better than for example crysis using tesselation engines done the proper way and at the same time make it more playable and smooth free seems to good to be true.

Question arise as, is it doable on current hardware with tesselation as in Fermi 480 or do we need another set of development for it to happen?
AMD at current cant do it up for the 6800 series, and the question is if they are planning to upgrade their 6900 series with Fermi like tesselation adaptions?

 

[Lonbjerg]

It makes AMD's demo look like the works of smurfs:
http://www.youtube.com/watch?v=vjtxy28FFN8

 

[Voo]

Anyone knows the specs of the used system (i.e. how many GPUs)? Just so that we can get a feeling for how many iterations we'll need before being able to do that stuff on a single high level GPU.

 

[Genx87]

I got the impression the first one was probably SLI? The smoke demo one at least SLI or more, and the last demo was able to be done on a single Fermi.

 

[Scali]

Ok I get that, fps is what counts especially playing shooters which you might not do.

Well, this demo was running in realtime, so apparently this technology can be applied with playable framerates.

However, if I understand you right, we can add more detail and create a lush for example jungle, which is even better than for example crysis using tesselation engines done the proper way and at the same time make it more playable and smooth free seems to good to be true.

It might sound too good to be true, but it isn't.
By generating extra detail on the fly, you reduce memory overhead and at the same time, you can reduce the amount of work for various calculations, such as animation and physics.

Question arise as, is it doable on current hardware with tesselation as in Fermi 480 or do we need another set of development for it to happen?

This demo was done on a Fermi chip, I assume a 480.
Even if it's a SLI setup, remember that they're talking about 500x more detail than current games.
So if it was a 2 GPU setup, then a single GPU can still handle more than 250x more detail... in the case of a 4 GPU setup, then it's still more than 125x more detail.
So apparently we can do this with hardware available on the market today.
125x more geometry detail than Crysis is nothing short of amazing.

 

[blastingcap]

Why are people thinking in terms of fps and such? You really need to stop doing that.
This example shows 1.3 billion triangles on screen. You can't even fit that amount in memory in the first place. Your average vertex takes about 32 bytes in memory, at the very least (position, normal vector, texture coordinates, that's 3 + 3 + 2 = 8 float values, 8*4 bytes = 32 bytes).
A triangle takes 3 vertices.
So we would need 1.3b * 3 * 32 = 124 GB of memory to store this geometry.

As they say, this is about 500x(!) more than what games have so far (Crysis averages around 2M polys on screen at the highest detail settings).

Basically this is simply impossible to do without tessellation. What's the framerate for impossible?
And THAT is what AMD is worried about. Tessellation is the only way forward for more detail in games. And AMD cannot do this. They want developers to lower the tessellation factors, and use more video memory.
With a proper tessellation implementation in hardware, we can do this NOW, as nVidia demonstrates. This is what DX11 is supposed to do.

For the next year, AMD's fixed tessellation solution will have to hold the fort. It appears to be a "good enough for now" solution. And guess what? It *is* good enough for now.

Regardless of what AMD wants, the market DEMANDS games be able to run without DX11. DX11 remains a small minority of the market, according to Steam Hardware Surveys. More importantly, no consoles are DX11. Until next-gen consoles launch, games will continue to be built on a DX9 chassis with some DX10/11 effects thrown in. You will get your tessellation wish no sooner than the next-gen consoles' launches, by which time AMD will probably have fixed its tessellation woes completely.

 

[Scali]

For the next year, AMD's fixed tessellation solution will have to hold the fort. It appears to be a "good enough for now" solution. And guess what? It *is* good enough for now.

Depends on who you're talking to.
I've been wanting to do tessellation on this sort of scale for many years (on my 8800GTS I had been experimenting with a vertex setup and tessellation pipeline written in Cuda, but I still had to rely on videomemory as intermediate storage on a per-mesh basis).
Now I have the hardware to make it happen.

Regardless of what AMD wants, the market DEMANDS games be able to run without DX11. DX11 remains a small minority of the market, according to Steam Hardware Surveys. More importantly, no consoles are DX11. Until next-gen consoles launch, games will continue to be built on a DX9 chassis with some DX10/11 effects thrown in. You will get your tessellation wish no sooner than the next-gen consoles' launches, by which time AMD will probably have fixed its tessellation woes completely.

It's not my 'wish'. I'm a developer. I can write this code today. I don't have to wait for it. I'll just leave you end-users playing with 5-year old technology and pretending that AMD's hardware is good DX11 hardware. As I said, ignorance is bliss.

 

[blastingcap]

Depends on who you're talking to.
I've been wanting to do tessellation on this sort of scale for many years (on my 8800GTS I had been experimenting with a vertex setup and tessellation pipeline written in Cuda, but I still had to rely on videomemory as intermediate storage on a per-mesh basis).
Now I have the hardware to make it happen.

It's not my 'wish'. I'm a developer. I can write this code today. I don't have to wait for it. I'll just leave you end-users playing with 5-year old technology and pretending that AMD's hardware is good DX11 hardware. As I said, ignorance is bliss.

Okay, then, we end-users will continue playing games using 5-year old technology. In the meantime, you will be playing with yourself. Unless of course you plan to make the first DX11-exclusive game or something and sell it on Steam.

Btw, NV's midrange parts are not *that* far ahead of Barts tessellation right now, according to tessellation benchmarks. You make it sound like an order of magnitude or something. All things considered, for the same price, I'd rather have a 6850 than the GTX460-768 I just bought.

 

[Voo]

This demo was done on a Fermi chip, I assume a 480.
Even if it's a SLI setup, remember that they're talking about 500x more detail than current games.

Actually they say "fully utilize several fermi GPUs" which is at least more than one, but other than that rather unspecific (although from the wording hints to more than 2 imho), so that'd be rather interesting.

 

[flopper]

Well, this demo was running in realtime, so apparently this technology can be applied with playable framerates.



It might sound too good to be true, but it isn't.
By generating extra detail on the fly, you reduce memory overhead and at the same time, you can reduce the amount of work for various calculations, such as animation and physics.

Did some read up as it havent been presented in any clear way why tesselation (tess) offers us benefits that cant be done with other rendering techniques.
One benefit as noted is to have instead of a 5.5mb texture in memory you are down to 130k...(did read on elite forum)
Which then makes a lot more sense towards eyecandy due if loosing the texture limts, we end up having a rich blooming game world that even with todays current hardware would be run not just fast but smooth.

So, the developer can make for example a MMOG where the world is rendered with tess and one art assest insetad of multiple oens thanks to adaptive tess which then saves time developing the game which can be spent on optimizing ther engine or other stuff.

Dont that also mean that the new engine for example BF3 (frostbite 2) will run with such mesh implementations from adaptive tess?
If so, that would be facinating to see.

 

[Idontcare]

In the meantime, you will be playing with yourself.

Uhm, all of us here are kinda geeky borderline nerdy types, right?

So aren't we all, kinda, ya know..."guilty as charged" at some point or another?

:$

 

[Gloomy]

Not me, I have dashing good looks and a silver tongue.:thumbsup:

 

[blastingcap]

Uhm, all of us here are kinda geeky borderline nerdy types, right?

So aren't we all, kinda, ya know..."guilty as charged" at some point or another?

:$

That wasn't the kind of "game" I was talking about, but ... um... guilty as charged, here. But only because my gf is currently so far away thanks to her job. We joke that she's a nympho when I'm around because she wants it like 1-2 times/day and I don't.

 

[Idontcare]

I'll just leave you end-users playing with 5-year old technology and pretending that AMD's hardware is good DX11 hardware. As I said, ignorance is bliss.

Scali wouldn't you agree though that in some cases "good enough" is, well, good enough?

I don't need a lambo to get from point A to point B even if the path separating A and B is an autobahn.

I do need a vehicle that meets a set of minimum operating characteristics (most notably a minimum max-speed).

So if the hardware DOES tessellation, then the amount it is doing is going to be up to me as I set my screen resolution, enable/disable eye-candy, as well as determine just how much of a slideshow I am personally willing to tolerate.

"Good enough" is subjective but the hardware must be capable of doing it.

AMD's cards are DX11 capable, unlike Intel's. No amount of "good enough" is feasible with Intel's graphics if the bare-minimum entry requirement is DX11/tessellation.

 

[blastingcap]

Not me, I have dashing good looks and a silver tongue.:thumbsup:

I've been called the Man with the Golden Tongue before. ()

 

[Lonbjerg]

Okay, then, we end-users will continue playing games using 5-year old technology. In the meantime, you will be playing with yourself. Unless of course you plan to make the first DX11-exclusive game or something and sell it on Steam.

Btw, NV's midrange parts are not *that* far ahead of Barts tessellation right now, according to tessellation benchmarks. You make it sound like an order of magnitude or something. All things considered, for the same price, I'd rather have a 6850 than the GTX460-768 I just bought.

If you really wan progress to slow, by all means...stick to consoles?
But that is hardly the point here, now is it?

And how far ahead tthe GTX460 is, depends on the level of tessellation, didn't you pay any attention in the other thread?

 

[Idontcare]

Not me, I have dashing good looks and a silver tongue.:thumbsup:

That wasn't the kind of "game" I was talking about, but ... um... guilty as charged, here. But only because my gf is currently so far away thanks to her job. We jokes that she's a nympho when I'm around because she wants it like 1-2 times/day and I don't.

Why I haven't a clue what you guys are referring to

:hmm:

:sneaky:

 

[blastingcap]

If you really wan progress to slow, by all means...stick to consoles?
But that is hardly the point here, now is it?

And how far ahead tthe GTX460 is, depends on the level of tessellation, didn't you pay any attention in the other thread?

I'm on your side here. I even bought a GTX460 for crying out loud (and yes I saw the other thread and no AMD wasn't THAT far behind with Barts. But I'm also a realist. Crytek gave up on the small PC gaming market (which is even smaller if you look at the DX11-capable portion of it (the OS and GPU must both be DX11 capable, plus a CPU that is capable of running games). Consoles are the 800 pound gorillas in the market. They dictate what we see available to a large degree.

 

[Lonbjerg]

I'm on your side here. I even bought a GTX460 for crying out loud (and yes I saw the other thread and no AMD wasn't THAT far behind with Barts. But I'm also a realist. Crytek gave up on the small PC gaming market (which is even smaller if you look at the DX11-capable portion of it (the OS and GPU must both be DX11 capable, plus a CPU that is capable of running games). Consoles are the 800 pound gorillas in the market. They dictate what we see available to a large degree.

At <16x tessellation...DX11 goes up to 64x...
(Infact AMD's tessellation performance tanks at already 11x...)

 

[Scali]

Did some read up as it havent been presented in any clear way why tesselation (tess) offers us benefits that cant be done with other rendering techniques.

I posted a link to the article on MSDN, explaining how and why tessellation gives you benefits you can't get otherwise, in one of the other threads about tessellation.
In short you reduce the geometry in memory, and also the amount of processing required on geometry (eg animation, interaction such as physics etc).
There basically is no way to get geometry this detailed without tessellation. Your memory is not large enough, and you won't have enough processing power to process everything at this high detail.

 

[Scali]

Scali wouldn't you agree though that in some cases "good enough" is, well, good enough?

Yes, but in this case AMD's stuff really is not "good enough".
The benefits of tessellation as shown in this video (incredibly high detail, and good performance) are out of reach of AMD's hardware.
Both performance gains and quality gains are very limited on AMD's hardware, if they exist at all (in some cases performance only goes down from tessellation, not up).

As I said before, it's pretty much the same as the geometry shader. A nice idea, but the hardware only got slower when you used the geometry shader, so you just went back to bruteforcing everything through the vertex shaders instead. Supporting a feature is useless when you cannot apply the feature in the way it was meant.

Think about it... see how low-detail the geometry in this demo is?
If you design your game this way, you NEED to handle a large tessellation factor. Otherwise you'll end up with lower detail than current games.
And at those tessellation factors, AMD's hardware will choke. So you won't get playable framerates.
So there are three options:
1) AMD cards will use tessellation, and get playable framerates, but at lower quality than current games.
2) AMD cards will use tessellation, and get good quality, but won't get playable framerates.
3) Games will have to be designed to not make optimal use of tessellation, so nVidia users cannot benefit from the fact that their hardware can handle such high quality at playable framerates either.

AMD wants to push for 3), which I find unacceptable.