"Yes, but how many polygons?" An artist blog entry with interesting numbers

[Inuhanyou]

Sorry if this has been asked before, but what do we think typical character models on the PS4/720 will be pushing? 50,000? 100,000? Or because of advanced texturing techniques and mapping, polygons aren't that big a deal anymore. Oh, and are we going to see tessellation play a bigger role?

Just to tantalize you, Naughty Dog has a job listing for someone who can 'render 1 million+ polygon characters for game ready applications'

http://www.gamerevolution.com/news/naughty-dog-pushes-for-one-million-polygons-10641

For comparison, something like Agni had 5 to 10 million poly's per frame,

http://www.finalfantasy-fxn.net/ind...-Philosophy-and-Luminous-Engine-Details--1384

i'm not sure about how many poly's the characters actually were though.

 

[3dcgi]

The quote from the job posting is "This digital sculptor will be responsible for making million poly models game ready". It doesn't say anything about the final poly count of the models.

 

[Inuhanyou]

That job listing is all we have to go on right now. So i posted it. I never insinuated anything about whether they would be scaled down for the final product, scaled up, and more importantly, we don't have any indication either way ourselves.

That's the beauty of speculation and anticipation though.

 

[lwill]

That job listing is all we have to go on right now. So i posted it. I never insinuated anything about whether they would be scaled down for the final product, scaled up, and more importantly, we don't have any indication either way ourselves.

That's the beauty of speculation and anticipation though.

Well, both Durango and PS4 GPUs are clocked at 1.6 billion polygon/sec, which is a little more than 3x the 360's. Shaders, tessellation, and lighting may be focused on more than current-gen, and the environment may need geometry more than the characters models will at such a higher level. I don't think we may see character polygon count dramatically higher than they are now, but it is best if an actual developer clarify on that point.

 

[scently]

Well, both Durango and PS4 GPUs are clocked at 1.6 billion polygon/sec, which is a little more than 3x the 360's. Shaders, tessellation, and lighting may be focused on more than current-gen, and the environment may need geometry more than the characters models will at such a higher level. I don't think we may see character polygon count dramatically higher than they are now, but it is best if an actual developer clarify on that point.

Yes but those figures does not take tessellation into consideration. So even if you keep the polygon count of this generation, the fact that these consoles can do tessellation means that the ingame model can end up looking much smoother as dynamic tessellation is applied on them. Also these systems are on a more modern and efficient architecture so they will be reaching their theoretical polygon/vertices performance better than the current generation.

 

[Laa-Yosh]

3dcgi is right, there's nothing about million polygon ingame models, it's just the current and well known pipe line - model/sculpt very highres meshes, build lowpoly versions, extract normal maps. Don't get yourself hyped up on a stupid misunderstanding.

 

[milk]

That 1.6 bilion polys/second is regarding total number of on-screen rendered polys, or total scene polys including the ones that end up culled?

By the way, we are aproaching micropolygon rendering at this rate aren't we? 1.6 bilion in 60fps means about 2 milion polys right? That's the amount of pixels on a 1080p screen, so on a best case scenario where only visible polys are rendered, and there's no transparent geometry (maybe even doable on a 2.5d game like trine or lbp) you could do a pseudo-reyes renderer in real time. That or I got my math wrong or I'm missed something.

 

[upnorthsox]

That 1.6 bilion polys/second is regarding total number of on-screen rendered polys, or total scene polys including the ones that end up culled?

By the way, we are aproaching micropolygon rendering at this rate aren't we? 1.6 bilion in 60fps means about 2 milion polys right? That's the amount of pixels on a 1080p screen, so on a best case scenario where only visible polys are rendered, and there's no transparent geometry (maybe even doable on a 2.5d game like trine or lbp) you could do a pseudo-reyes renderer in real time. That or I got my math wrong or I'm missed something.

20 mil.

 

[milk]

So that means you have 10 polys per pixel, on a 1080p60 game. Double of that at 30fps and even more if you are rendering at a lower res. Of course real world scenarios wouldn't be that performat, but some that is still a lot isn't it? Isn't that going to require some rethinking of current rendering pipelines to get the most out of a next gen engine?

 

[3dcgi]

That 1.6 bilion polys/second is regarding total number of on-screen rendered polys, or total scene polys including the ones that end up culled?

By the way, we are aproaching micropolygon rendering at this rate aren't we? 1.6 bilion in 60fps means about 2 milion polys right? That's the amount of pixels on a 1080p screen, so on a best case scenario where only visible polys are rendered, and there's no transparent geometry (maybe even doable on a 2.5d game like trine or lbp) you could do a pseudo-reyes renderer in real time. That or I got my math wrong or I'm missed something.

Games spend a lot of time rendering things that aren't visible in the final image as polygons per se. Shadow maps for example.

 

[ERP]

It's a number that describe the peak performance of one piece of a pipeline, practically the only way you'll see anything close to that number is if the vertex shader is trivial and all the polygons are blackface culled.
Go read the ATI docs and look at what the restrictions are on that number.

 

[lwill]

It's a number that describe the peak performance of one piece of a pipeline, practically the only way you'll see anything close to that number is if the vertex shader is trivial and all the polygons are blackface culled.
Go read the ATI docs and look at what the restrictions are on that number.

So the ratio between peak performance vs real world performance will be similar to what it has been for the current-gen consoles (like 20-40% of peak performance)?

 

[Laa-Yosh]

I wouldn't be surprised if the ratio would go down a little further. More complex shaders on one hand, loss of fragment efficiency because of smaller triangles, and so on.

Also, characters for example are made up of vertices that take longer to process, because of skinning and possibly simple cloth/hair physics, and blendshapes and such. That's another factor that reduces the real world results.

Again, millions of polygons per character in this generation are very unlikely, even if some game ends up implementing tessellation and displacement on characters.

 

[Shifty Geezer]

Can we assume a pretty universal tessellation being applied to smooth objects, just doubling up points as a form of smoothing except in specific cases (faces) where definition needs to be preserved? Things like shoulders have been angular forever, and it'd be nice to think these'll get automagically smoothed out a little which I expect interpolated points to do.

 

[Laa-Yosh]

I don't think so - blindly tessellating would ruin at least as much on a model as it would fix...

But the systems should be powerful enough to have the necessary definition in the lowres mesh already.
For example do you have any such problems with the characters in Last of Us? That level of detail should pretty much become the standard on the next gen...

Of course there'll always be some games with too tight budgets where there isn't enough time/money/people to fix everything but then those games wouldn't fix it using tessellation either.

 

[Shifty Geezer]

Tessellation instead of high-res models will reduce memory use and BW (RAM and IO). Just finding an example pic, this Borderlands 2 shot shows the pointy shoulders I'm used to seeing. To my ignorant eye, it seems a simple job to smooth that out by adding interpolated triangles. If not the case, tessellation is sounding increasingly limited! It'll add lumps and bumps to trees and rocks. Whoop-de-doo.

 

[lwill]

I wouldn't be surprised if the ratio would go down a little further. More complex shaders on one hand, loss of fragment efficiency because of smaller triangles, and so on.

Also, characters for example are made up of vertices that take longer to process, because of skinning and possibly simple cloth/hair physics, and blendshapes and such. That's another factor that reduces the real world results.

Again, millions of polygons per character in this generation are very unlikely, even if some game ends up implementing tessellation and displacement on characters.

Oh, I see. It is as I thought, though I wasn't sure on how tessellation would affect polygon rendering.

I don't think so - blindly tessellating would ruin at least as much on a model as it would fix...

But the systems should be powerful enough to have the necessary definition in the lowres mesh already.
For example do you have any such problems with the characters in Last of Us? That level of detail should pretty much become the standard on the next gen...

Of course there'll always be some games with too tight budgets where there isn't enough time/money/people to fix everything but then those games wouldn't fix it using tessellation either.

Thank you for clarifying this. I honestly thought tessellation would be a bit simpler to implement, though that does explain it is still not in common use.

What polygon count range are you expecting for environments and character models for next-gen games?

 

[yewyew]

Tessellation instead of high-res models will reduce memory use and BW (RAM and IO). Just finding an example pic, this Borderlands 2 shot shows the pointy shoulders I'm used to seeing. To my ignorant eye, it seems a simple job to smooth that out by adding interpolated triangles. If not the case, tessellation is sounding increasingly limited! It'll add lumps and bumps to trees and rocks. Whoop-de-doo.

Tessellation should never be looked at as "the next big thing" or a marketing point in API's or whatever. It's all down to good coding in the end. Hell, guys like Naughty Dog have been using tessellation in their LOD's since Crash Team Racing on PS1, did you notice it then?

http://www.gamasutra.com/view/feature/131394/postmortem_naughty_dogs_jak_and_.php?print=1

 

[Shifty Geezer]

LOD tessellation/mesh reduction has been around since DX7 at least. The new thing featuring in hardware on the GPU is on-the-fly triangle creation, so constructive tessellation, and it is something new. It's just something I don't fully understand regards what it brings to modern rendering, other than the displaced tree examples. Effectively it seems more about adding geometry complex by combining extended art assets (displacement maps and high-res models) and not procedurally smoothing or rounding geometry on any level.

 

[DRS]

I don't think so - blindly tessellating would ruin at least as much on a model as it would fix...

It doesn't have to be blindly, it can use the vertex normals to estimate a curve, right? That's how I would implement such a thing at least (and use additional heightmaps).