[SemiAccurate] Nvidia's Fermi GTX480 is broken and unfixable

[thilanliyan]

people are starting to give less importance to tesselation since the data of how good Fermi is at it.

Where are people doing that?

 

[NoQuarter]

Where are people doing that?

I know, I haven't seen any talk like that. I hope Nvidia does tessellation well so more game devs will put it in and I can see it on the DX11 card I already got.

 

[dguy6789]

Where are people doing that?

He thinks AMD fans will say tessellation isn't important because AMD isn't as good at it just the way Nvidia fans said DX11 isn't important for the last 6 months.

I don't think this is going to happen. Tessellation is amazing and has great potential to make games more realistic looking. It would be a little silly to try to downplay its importance.

 

[Schmide]

Tesselation on GF100 is performed in the polymorph engines. Each cluster of 32 shaders has 1 polymorph engine. Total of 16 polymorph engines per GF100.

So if a 128 shader SKU is released, it would have 4 polymorph engines, 256/8, 384/12.
It appears tesselation is scalable the higher you go in GPU rank.

It just seems the polymorph engine is far from a fixed function pipeline, which is probably near what you would call the ATI unit. Other than a dramatic change in primitive (F16, RGB8, etc) it could probably be reprogrammed to handle many future APIs.

It looks to me like they extracted specific function towards a general purpose unit (almost CPU like), to perform the needed tasks. This abstraction though extremely scalable is not without its complexity problems, out of order execution, data dependencies etc. I often wonder what decisions are made in Fermi with regards to how much processing power to dedicate to each rendering task.

 

[cbn]

He thinks AMD fans will say tessellation isn't important because AMD isn't as good at it just the way Nvidia fans said DX11 isn't important for the last 6 months.

I don't think this is going to happen. Tessellation is amazing and has great potential to make games more realistic looking. It would be a little silly to try to downplay its importance.

In the Unigine benchmark ATI hardware takes a big hit from turning tessellation on in DX11.

I can't remember the exact numbers but the delta was quite large. I want to say FPS came close to dropping by half (but I am not sure).

Tessellation is exciting technology, I just hope ATI can deliver on that in the future rather than just providing high average FPS numbers. I mean it is usually the minimum FPS most of us are worried about right?

 

[dguy6789]

It's not really clear if the Unigine benchmark reflects real world performance yet. The level of tessellation used in it might be unreasonably heavy compared to what will be seen in games.

If you run the benchmark without tessellation on, it looks really ugly. One example is that with tessellation off, the benchmark doesn't even render stairs, something that has been done in games without the aid of tessellation for a decade at least. It seems to me that the benchmark is using tessellation for virtually everything rather than using it only on objects that would benefit from or need the high polygon count.

ATI has had some form of tessellation in every graphics card they have made since 2001, I imagine they would have some idea of what they are doing. It's too early to draw conclusions about the 5870's tessellation abilities.

 

[Keysplayr]

Hmmmmf.. Ok, I take it back. Sounds like most of us are pretty excited about it.

 

[tviceman]

Hmmmmf.. Ok, I take it back. Sounds like most of us are pretty excited about it.

I am really excited! When Metro 2033 announced all it's high end graphical features, including physx and tessellation, it shot to the top of my want-to-buy list.

 

[Keysplayr]

It just seems the polymorph engine is far from a fixed function pipeline, which is probably near what you would call the ATI unit. Other than a dramatic change in primitive (F16, RGB8, etc) it could probably be reprogrammed to handle many future APIs.

It looks to me like they extracted specific function towards a general purpose unit (almost CPU like), to perform the needed tasks. This abstraction though extremely scalable is not without its complexity problems, out of order execution, data dependencies etc. I often wonder what decisions are made in Fermi with regards to how much processing power to dedicate to each rendering task.

In any event, GF100 has dedicated hardware outside of it's shaders to perform tesselation. I assumed that would be all that anyone was concerned with. For example, if tesselation was run on the shaders, would it rob the GPU of some performance while doing it. What other concerns would there be?

 

[Schmide]

In any event, GF100 has dedicated hardware outside of it's shaders to perform tesselation. I assumed that would be all that anyone was concerned with. For example, if tesselation was run on the shaders, would it rob the GPU of some performance while doing it. What other concerns would there be?

I really have no idea how many inter dependencies are created in Fermi's parallel setup geometry units. Assuming the shaders are dependent on the geometry engine and tessellation data has to go through some global ordering, I can imagine some shaders being stalled by it.

That being said, there probably isn't a geometry pipeline in existence that isn't stalled by some operation or another.

 

[thilanliyan]

He thinks AMD fans will say tessellation isn't important because AMD isn't as good at it just the way Nvidia fans said DX11 isn't important for the last 6 months.

I don't think this is going to happen. Tessellation is amazing and has great potential to make games more realistic looking. It would be a little silly to try to downplay its importance.

I know he thinks that but I haven't seen it turn into that yet. I'm extremely excited about Tesselation...thankfully it is a DX standard (too bad it was not included in DX10) and will gain ground quickly I hope. Like another person mentioned, Metro 2033 looks pretty good for some of us graphics whores and I could care less whether it is being run on an nV or ATI card.

 

[cbn]

Hmmmmf.. Ok, I take it back. Sounds like most of us are pretty excited about it.

Just wondering.....do you think tessellation could improve Nvidia's 3D vision effect?

Would these two technologies be synergistic with one another? Or are we just talking an additve effect?

 

[Itchrelief]

Just wondering.....do you think tessellation could improve Nvidia's 3D vision effect?

Would these two technologies be synergistic with one another? Or are we just talking an additve effect?

What do you mean, like if they exaggerated all the tesselation on the z-axis to make it jump out more with 3d glasses?

I would figure it would be just as synergistic as most any other technology.

 

[cbn]

What do you mean, like if they exaggerated all the tesselation on the z-axis to make it jump out more with 3d glasses?

Yes, that is what I am talking about. Thank you for putting that into more accurate terms.

 

[cbn]

Speaking of tessellation,

Does anyone think this could be a more important graphics setting than AA?

With pixels becoming smaller I have to wonder if more people would rather sacrifice AA performance for more tessellation.

 

[Fox5]

Speaking of tessellation,

Does anyone think this could be a more important graphics setting than AA?

With pixels becoming smaller I have to wonder if more people would rather sacrifice AA performance for more tessellation.

I would think with more polygons, AA is more essential?

 

[jvroig]

In any event, GF100 has dedicated hardware outside of it's shaders to perform tesselation. I assumed that would be all that anyone was concerned with. For example, if tesselation was run on the shaders, would it rob the GPU of some performance while doing it. What other concerns would there be?

Right, thanks. I don't see any other concern outside of "would it rob the GPU of performance", hence the desire for dedicated hardware for it.

I wonder though, if ATI is using dedicated hardware for it, why then does it take a huge hit in the Unigine benchmark that Bottleneck refers to? And would it theoretically have the same effect with Fermi cards since both seem to have dedicated hardware for tessellation?

 

[Schmide]

My spin on modern rendering with tessellation.

Traditionally rendering systems have been bottlenecked by the shaders. As much data as you could setup on the CPU and push through the bus could be handled by the transform and lighting engine with more than enough throughput for final rasterization. What tessellation has done is eased the burden on the CPU and bus causing the feeding gap of T&L engine to be filled; thus, possibly becoming a bottleneck, and consequently justifying the need for more rasterization (shader/texture) units. Antialiasing/anisotropic filtering only burden these rasterization units more.

 

[Keysplayr]

Right, thanks. I don't see any other concern outside of "would it rob the GPU of performance", hence the desire for dedicated hardware for it.

I wonder though, if ATI is using dedicated hardware for it, why then does it take a huge hit in the Unigine benchmark that Bottleneck refers to? And would it theoretically have the same effect with Fermi cards since both seem to have dedicated hardware for tessellation?

Yeah tesselation does add a lot more triangles to the mix. So I guess more of a workload on the GPU after the geometry setup is done. Or am I not on the right track? Anyone?

 

[cbn]

I wonder though, if ATI is using dedicated hardware for it, why then does it take a huge hit in the Unigine benchmark that Bottleneck refers to?

http://www.hardocp.com/article/2009/11/06/unigine_heaven_benchmark_dx11_tessellation/

Here is the benchmark I was talking about earlier (I finally found it). Notice AA is not being used.

 

[blanketyblank]

http://www.hardocp.com/article/2009/11/06/unigine_heaven_benchmark_dx11_tessellation/

Here is the benchmark I was talking about earlier (I finally found it). Notice AA is not being used.

http://www.xbitlabs.com/articles/video/display/radeon-hd5770-hd5750-crossfirex_6.html#sect3

One of the reasons why I went with a 5770 is it seems to scale really well with tesselation enabled in ungine, 87% with AA. And with 8xAA it is almost playable at 25 fps, 43 without.
If it's seperate hardware on the ATI cards and it scales so well in crossfire, it makes me wonder if they can't make a dedicated tesselation card in the future if they wanted to.

 

[gorobei]

Originally Posted by Keysplayr
people are starting to give less importance to tesselation since the data of how good Fermi is at it.

He thinks AMD fans will say tessellation isn't important because AMD isn't as good at it just the way Nvidia fans said DX11 isn't important for the last 6 months.

I don't think this is going to happen. Tessellation is amazing and has great potential to make games more realistic looking. It would be a little silly to try to downplay its importance.

no one is diminishing the importance of tess, just the implementation of gf100 tess. Most of those who read the reviews of the gf100 on anand understood that with each shader module/cluster equipped with a polymorph engine and dynamically shifting clusters over to tessellating as needed, could result in less shaders/clusters being available for normal aa/af/lighting functions. As long as gf100 can run Pmorph functions each cycle alongside it's SM, it's a nice bump in performance. BUT if gf100 loses SM function because the Pmorph is being used for tess, then gf100 isn't as impressive. Real world games unlike the unigine, put more demands on the gpu. We would need to see a AvP benchmark in order to know whether fermi would really get 2x the performance over 5870 in dx 11 tess enabled games.(which most people acknowledge won't be the case)

My spin on modern rendering with tessellation.

Traditionally rendering systems have been bottlenecked by the shaders. As much data as you could setup on the CPU and push through the bus could be handled by the transform and lighting engine with more than enough throughput for final rasterization. What tessellation has done is eased the burden on the CPU and bus causing the feeding gap of T&L engine to be filled; thus, possibly becoming a bottleneck, and consequently justifying the need for more rasterization (shader/texture) units. Antialiasing/anisotropic filtering only burden these rasterization units more.

yes and no. it depends on how the model and subdivision surface/tessellation algorithm is setup. Regardless of what algorithm dx11/game-engine is using, the models being sent from the cpu to the gpu are normally going to be roughly the same number of polys we're using now for DX9. It's just the nature of the modelling and of limit surfaces. It's only extreme situations like the unigine demo that the models were radically simplified to show what tessellation can do that the polycount was far less (like the stairs being a single flat poly). Any normal game would just build regular stairs since the collision/dynamics LOD would need that info anyways and just use less tessellation on those properly built stairs. By going with a flat plane, the unigine demo requires using full tess to just to get a simple collision object. That maybe fine when we want to see individual droplets of water running through the crags of each rock in the stair, but we're probably years away from that. Extreme reliance on tess and displacement also results in texture/UV issues. If you look closely at unigine, there are areas that show massive texture stretching. It isnt a big deal as far as just looking good, but when it comes to decal (bullet hits/etc) placement there will be problems.

the real advancement will be in instancing, which was available in dx10. When you get a game using instancing to it's full extent, game visuals will be far more interesting.

 

[cbn]

http://www.xbitlabs.com/articles/video/display/radeon-hd5770-hd5750-crossfirex_6.html#sect3

One of the reasons why I went with a 5770 is it seems to scale really well with tesselation enabled in ungine, 87% with AA. And with 8xAA it is almost playable at 25 fps, 43 without.
If it's seperate hardware on the ATI cards and it scales so well in crossfire, it makes me wonder if they can't make a dedicated tesselation card in the future if they wanted to.

Where does it say Tesselation was enabled in the Unigine heaven demo? Did they run "Heaven" without Tessellation also?

 

[GodisanAtheist]

Where does it say Tesselation was enabled in the Unigine heaven demo? Did they run "Heaven" without Tessellation also?

According to the heavenly demo from Unigine, the CrossFireX configurations are faster than their single-GPU opponents Radeon HD 5870 and HD 5850. The two junior Radeon HD 5750 cards are as fast together as the single Radeon HD 5870. The high results of the GeForce GTX 260 and Radeon HD 4770 (in ordinary mode) are due to the lack of DirectX 11 support whereas the cards based on AMD’s new GPUs all work in DirectX 11 mode.

-The fact that they point out that the GTX 260 & 4770's results are skewed by the lack of DX11 implies that everything else was run in DX11.

 

[3DVagabond]

I've seen the reverse of tessellation, removing polies, done by software before. It's always totally screwed the smoothing of the model. Does anyone know how tessellation deals with smoothing groups or smoothing (phong) angles? Does it rely on the information from the model? does it just do it by some predetermined angle? Anyone know? I'm curious because I do game models and this type of "feature" has always screwed up the models more than anything else. Also, models in games are made up of triangles. Doing SubD on a triangulated model is a no-no. Does anyone know how this is dealt with?