Tesselation review by xbitlabs

[Scali]

You contradict yourself

No I don't.
Unigine is a game engine. Heaven is a synthetic benchmark.

It seems it's not a weakness. Yes AMD tessellation unit is not as strong as NV one but I would say it is good enough for any reasonable in game tessellation (that improves image quality and does not force a work on invisible areas).

We have yet to see any reasonable in-game tessellation, so it's too early to call that.
I'd say that Unigine's Heaven benchmark is the most realistic use of tessellation we have so far (as in, proper low-poly geometry and reasonable amplification factor).
I think the upcoming 3DMark11 later this year will show us how tessellation is done properly:
http://www.youtube.com/watch?feature=iv&fmt=37&v=xEQrPZntkU8

 

[evolucion8]

In simple words, Unigine's Engine (Doesn't matter if its synthettic or not because its a game engine which can be used for a game), uses Tessellation in a very innefficient way, hopefully Futuremark will show how it's done if the don't dwell too much in architecture's preference.

 

[Zstream]

Depends on which benchmark you take.
Obviously with programmable hardware, it all depends on how the functionality is implemented by the developer.
Stone Giant and Heaven show exactly what nVidia had in mind: There is a pretty linear decrease in performance going from low to high tessellation settings.
The Radeons seem to have 'reached their fill' at a certain point and drop off almost exponentially from there on.

Heaven relies almost completely on tessellation, where turning tessellation off results in 'wrong' geometry, because too much detail is missing.

But that's not a realistic scenario for games. If you disable tessellation in games, you still want things to look correctly. I think this results in the 'flaw' that games with tessellation are based on the non-tessellated geometry, rather than simplified geometry (which is how tessellation should really be used... reduce detail).

Um, the reason why Nvidia does better is due to the hardware dedicated to it. Nvidia dedicated a large fraction of the chip to Tessellation. In fact if you look at the infastructure, ATI is doing much more with less.

 

[RussianSensation]

More likely than not, by the time Extreme Tessellation as used in Unigine is around, GTX480 and 5870 will both be replaced by next gen cards.

What we can see though is that Fermi is a bit more forward looking as it was able to outperform the 5870 without Tessellation turned on in AvP, STALKER: CoP and Metro 2033 and Dirt 2. This means that ATI will need to work just a little bit harder on their next card to catch up in DX11 performance (irrespective of Tessellation).

 

[evolucion8]

More likely than not, by the time Extreme Tessellation as used in Unigine is around, GTX480 and 5870 will both be replaced by next gen cards.

What we can see though is that Fermi is a bit more forward looking as it was able to outperform the 5870 without Tessellation turned on in AvP, STALKER: CoP and Metro 2033 and Dirt 2. This means that ATI will need to work just a little bit harder on their next card to catch up in DX11 performance (irrespective of Tessellation).

Fermi you mean the GTX 480 right? Because the GTX 470 can almost go toe to toe with the HD 5870, and that's without the full 512 stream processors, if nVidia launch it, AMD will have to do something, the HD 5970 will remain being the fastest card, but would not be by much and the price difference will force AMD to lower its price which will be good for us the consumers.

 

[Wreckage]

Wow if the Heaven benchmark is an indication of future tessellation performance look at how well the 460 did against the 5870 The Stone Giant demo seems to confirm this.

Santa Clara based Nvidia Corporation hails Fermi chips as a first family of graphics cards that were designed with hardware tessellation technology as its cornerstone and Heaven Benchmark can serve as a testament to this claim. GeForce GTX 480 and GeForce GTX 460 indeed fair a bit better than their competitors from AMD camp.

Excellent review. Hopefully with more mature drivers the 460 will climb even higher. :thumbsup:

 

[Xarick]

The problem with heaven is it is not a realistic picture of gaming. That is just tess. What that shows is Nvidia has a more efficient tess solution, but none of them are gonna be enough for games to use high levels of tess effectively. We are still a few gens away.

 

[evolucion8]

Yup, that's correct. Tessellation performance is only an angle, other stuff like the weaker texture performance the GTX 4x0 series has compared to Cypress, lower pixel fillrate, ROP performance and shader performance its all a plus in the final rendering of the game, all that sum of the factors which will determine which offers the best performance in games. For know on, we can state that by the time that DX11 is widely used, none of the current solutions will be able to play it properly and newer and more powerful solutions like Souther Islands or N.I. should be able to offer the performance needed. The funny thing is that regardless of the powerful Polymorph engines in the Fermi cards, only offered 10 percent better performance across tittles in average in the best case scenario, a die so big like Fermi barely competing against the much smaller Cypress, interesting. More than 10 percent performance boost is needed for the GTX 460 to catch up with the faster HD 5870 Not even close in Metro 2033 with favors heavily nVidia.


"In terms of pure performance drops in regards of AMD and Nvidia architecture it is really hard to judge which architecture tessellation algorithms prefer at the moment. In case of S.T.A.L.K.E.R.: Call of Pripyat, Aliens vs. Predator and Colin McRae: Dirt 2 the overall drop in fps is either very small or remains the same across the board. At the same time, when tessellation workload increases, Nvidia’s Fermi products demonstrate a slightly smaller tendency to lose frames. Stone Giant and Heaven benchmarks, as well as Metro 2033 video game clearly prefer Santa Clara designed products. On average, Nvidia based products are from 5 to 10 percent more efficient in terms of hardware tessellation calculations. But you have to remember that both Stone Giant and Heaven are synthetic benchmarks without any video game implementation at the moment, and Metro 2033 is just one game, which is hardly a determinative factor."

 

[Scali]

In simple words, Unigine's Engine (Doesn't matter if its synthettic or not because its a game engine which can be used for a game), uses Tessellation in a very innefficient way, hopefully Futuremark will show how it's done if the don't dwell too much in architecture's preference.

Beyond3D has not provided any proof that it is in fact inefficient.
Namely, they have made no attempt to compare the cost of backface culling and rendering zero-area triangles against the cost of trying to elimiate these triangles higher up in the tessellation process.

Since the GeForce, backface culling and rendering zero-area triangles have been extremely cheap. So cheap that it was always better to just send them to the renderer, than to try and eliminate them earlier in the process.
I'm not sure if any of that has changed... or how expensive Heaven's tessellation step is anyway.

Aside from that, the *Engine* doesn't tell you how it does tessellation. The *programmable shaders* do.
So it's perfectly possible to make an engine based on Unigine that uses a different tessellation approach.

 

[Scali]

Um, the reason why Nvidia does better is due to the hardware dedicated to it.

Yes, but how is a synthetic benchmark at fault for demonstrating that nVidia's approach actually delivers better tessellation than ATi's approach in practice?

ATi simply under-designed their tessellator, making it pretty much useless in practice.
nVidia's tessellator may not be powerful enough yet either, but at least they're on the right track. They can just scale up their design.
ATi needs to completely redesign it, and build a parallel solution.

 

[Scali]

but none of them are gonna be enough for games to use high levels of tess effectively.

I don't see why not, really.
Unigine runs at reasonable framerates, right? You can play games at those framerates.
And if Beyond3D's claims are true, there is room to optimize performance even more.

 

[Scali]

other stuff like the weaker texture performance the GTX 4x0 series has compared to Cypress

The thing is... games use textures and pixel shading tricks to compensate for lack of geometry detail... bumpmaps, occlusion maps, parallax mapping etc.
If you start using tessellation, you can drop a few of these textures and per-pixel operations. So a fast tessellator can compensate for texture and pixel performance to a certain extent.

 

[Xarick]

I don't see why not, really.
Unigine runs at reasonable framerates, right? You can play games at those framerates.
And if Beyond3D's claims are true, there is room to optimize performance even more.

There is nothing going on. In a game your gonna have all kinds of stuff going on. Granted they would have to tone down the graphics considerably for a game to run nice with everything else, but then so would the tess drop considerably.

 

[Scali]

There is nothing going on. In a game your gonna have all kinds of stuff going on.

Things that would generally be done on the CPU, so they won't have much to do with the GPU and its handling of tessellation.

 

[Zstream]

Yes, but how is a synthetic benchmark at fault for demonstrating that nVidia's approach actually delivers better tessellation than ATi's approach in practice?

ATi simply under-designed their tessellator, making it pretty much useless in practice.
nVidia's tessellator may not be powerful enough yet either, but at least they're on the right track. They can just scale up their design.
ATi needs to completely redesign it, and build a parallel solution.

*sigh* Are you at all familar with how it works? ATI is getting a MUCH larger performance gain with less hardware used. Your statement that they need to go back to the drawing board is completely insane.

 

[Scali]

*sigh* Are you at all familar with how it works? ATI is getting a MUCH larger performance gain with less hardware used. Your statement that they need to go back to the drawing board is completely insane.

I'm an M.Sc. in computer graphics. I like to think I know how tessellation works, yes.
ATi's hardware is under-equipped for proper tessellation (think NURBS etc, Pixar style).
I don't really care if they get 'more performance gain' with 'less hardware' if the performance in the absolute sense is not good enough. Are you at all familiar with what I mean by that?
Benchmarks clearly demonstrate that there's a big bottleneck with large amplification factors on ATi hardware. You see an exponential dropoff in performance. nVidia's dropoff is linear, or even sub-linear.

 

[tviceman]

Slightly, but not completely, off topic: Civilization V will utilize tessellation for it's DX11 mode. Personally I think it's fantastic news! The industry seems to be adopting DX11 much more willingly than the largely ignored DX10.

 

[Grooveriding]

I'm an M.Sc. in computer graphics. I like to think I know how tessellation works, yes.
ATi's hardware is under-equipped for proper tessellation (think NURBS etc, Pixar style).
I don't really care if they get 'more performance gain' with 'less hardware' if the performance in the absolute sense is not good enough. Are you at all familiar with what I mean by that?
Benchmarks clearly demonstrate that there's a big bottleneck with large amplification factors on ATi hardware. You see an exponential dropoff in performance. nVidia's dropoff is linear, or even sub-linear.

I'm a Ph.D. in organic chemistry, but I don't presume to know how to design an oscillator. Get back to us when you're an engineer and can claim expertise.

 

[Scali]

I'm a Ph.D. in organic chemistry, but I don't presume to know how to design an oscillator. Get back to us when you're an engineer and can claim expertise.

Uhhh, I work as a software engineer (specialized in computer graphics obviously, given my education), and I've implemented my share of tessellation routines over the years.
I probably know more about the subject than anyone on this forum, so don't give me this crap.

 

[Grooveriding]

Uhhh, I work as a software engineer (specialized in computer graphics obviously, given my education), and I've implemented my share of tessellation routines over the years.
I probably know more about the subject than anyone on this forum, so don't give me this crap.

So as a software engineer you design hardware ? Or your work makes use of the hardware as a tool ?

 

[evolucion8]

Scali, no matter what AMD does, you are never sattisfied with it, in that review, it was proven that nVidia's not much faster in average than AMD with Tessellation, and by the time that the HD 5870 is unplayable with Tessellation, Fermi will be also unplayable (Except the GTX 480 which is not meant to compete with the HD 5870). Currently both enjoy great frame rate with games that uses Tessellation like Dirt 2, AvP, Metro 2033, STALKER COP, Unigine is a show of the worst case scenario which none of the cards used performs at playable frame rate, please, stop being so pro nVidia and show some neutrality.

 

[Scali]

So as a software engineer you design hardware ? Or your work makes use of the hardware as a tool ?

Why would I need to design hardware?
Does a formula 1 driver also need to be able to design a race car? Do race car engineers need to be formula 1-class drivers?

Heck, hardware engineers may be able to design a GPU, but I wouldn't want them to design a game engine.

I am the 'driver' of the hardware, and as such I have the best judgement of where potential bottlenecks are.
I don't need to know how the hardware works at a transistor level, as that is completely irrelevant.

 

[Scali]

Scali, no matter what AMD does, you are never sattisfied with it

Sure I am. I bought a Radeon 5770, didn't I?
But that was then, and this is now.

it was proven that nVidia's not much faster in average than AMD with Tessellation

It was proven that nVidia's performance has a linear dropoff characteristic, while AMD's performance has an exponential characteristic.

I don't understand why we even need to have this discussion, in the days of multicore CPUs and everything.
AMD has made a serial implementation, capable of handling only one triangle at a time. That's where the bottleneck is. nVidia has parallelized this step, and that is obviously the way of the future.

stop being so pro nVidia and show some neutrality.

I'm as neutral as can be when I say nVidia's tessellator is better.
I'm also as neutral as can be when I say that current games don't use tessellation the way it's meant to be used. I've referred to Pixar numerous times, and I will do so again.
Google Pixar Renderman REYES, and read up on how Pixar applies tessellation.
That's what we want in the future. Now, Pixar is completely neutral, right? They have nothing to do with either nVidia or AMD, agreed? Good.

 

[Zstream]

It was proven that nVidia's performance has a linear dropoff characteristic, while AMD's performance has an exponential characteristic.

I'm as neutral as can be when I say nVidia's tessellator is better.
I'm also as neutral as can be when I say that current games don't use tessellation the way it's meant to be used. I've referred to Pixar numerous times, and I will do so again.
Google Pixar Renderman REYES, and read up on how Pixar applies tessellation.
That's what we want in the future. Now, Pixar is completely neutral, right? They have nothing to do with either nVidia or AMD, agreed? Good.

Ok, let me explain it again.

FACT: ATI's implementation when taking into account the size of the tessellation hardware is better then Nvidia's.

FACT: Nvidia's cards are performing better then ATI's card due to hardware size and not software.

 

[Wreckage]

Ok, let me explain it again.

FACT: ATI's implementation when taking into account the size of the tessellation hardware is better then Nvidia's.

FACT: Nvidia's cards are performing better then ATI's card due to hardware size and not software.

How are these facts? Do we know how much of each chip is dedicated to tessellation? From everything I have read NVIDIA has a larger chip for GPGPU/CUDA support.

Can you back any of this up? :thumbsdown:

The only thing I see that is a fact (based on this article) is that NVIDIA stomps ATI when using tessellation. So much so that GTX460 can surpass the 5870. That's all that is important. In the end no gamer cares about the hardware, they care about how their games play and look.