[Extremetech] Nvidia’s GameWorks program usurps power from developers, end-users, AMD

[Lonbjerg]

That makes sense sontin, the Anandtech review mentioned that the performance gain in tessellation at higher factors over Tahiti corresponded to the memory bandwidth gain. If the tessellator is choking by having to go off-chip, this makes sense.



Because it's tessellating something you can't see using that 20% performance.

Are you sugesting that you cannot tell non-tesselated from tesselated?

 

[blastingcap]

Are you sugesting that you cannot tell non-tesselated from tesselated?

It can be hard to see the benefit of tessellation in games. DX9 (zero tessellation) isn't that much worse than DX11 (tessellation) depending on how much effort developers put into the DX9 version. So going from moderate tessellation to extreme tessellation would yield even smaller gains--almost imperceptible. And literally imperceptible if the tessellation is for off-screen or flat stuff like the infamous Crysis 2 stuff.

http://www.overclock.net/t/690441/pcgh-metro-2033-direct-x11-comparison-screenshots <-- NVIDIA sponsored game
http://www.pcper.com/news/General-Tech/Never-mind-barrier-we-even-tessellated-water-table <-- NVIDIA sponsored game
http://www.geforce.com/whats-new/gu...ty-graphics-breakdown-and-performance-guide#1 <-- NVIDIA sponsored game and this time NVIDIA no doubt stung by the reception of tessellation so far, probably made it a point to get the Batman:AC devs to deliberately make a larger difference between the DX9 and DX11 versions just to "prove" how big a difference tessellation can make... you can tell this was a priority for NVIDIA because they put it on their own site, geforce.com, to attempt to convince gamers that you should buy NVIDIA cards)

Businesses like multiplatform development to cut costs. Consoles use GCN. AFAIK the consoles don't get special tessellators. Thus I expect us to be limited by AMD's tessellation ability for most multiplatform games, except the ones that NV pays to be extra-tessellated on PC. This is like the DX11 situation where, because consoles were DX9 limited until recently, multiplatform games were pretty much all DX9 except for some that got tacked-on DX11 that often weren't worth the performance drop.

 

[sontin]

That's because character models are high poly models in games.
You see the benefit much better with low poly models or flat surfaces - for example the terrain with and without Tessellation in HAWX 2.

 

[Anarchist420]

Go back to P&N, you don't know what you're talking about.

if you know so much that i dont then why dont you correct what i said that you believe to be wrong?

This is satire, right?

nope

 

[blastingcap]

That's because character models are high poly models in games.
You see the benefit much better with low poly models or flat surfaces - for example the terrain with and without Tessellation in HAWX 2.

That's the problem though--nobody looks at the terrain most of the time. When you are in a dogfight, if your focus is on the terrain rather than the enemy, you're doing it wrong. Yes you get moderately prettier backgrounds if you are flying in peaceful territory but that's of limited use outside of taking screenshots. Look at the B:AC link. If your focus is on the brick wall rather than Batgirl's face (or whoever that woman is), you're fapping it wrong.

Part of the problem is that human beings' eyes are very blurry outside of a tiny cone. It takes a supercomputer (your brain) to process the raw signals sent from your eyes and make it acceptably sharp in the middle. As an example, focus on this word on the screen:



WORD



Chances are, at typical viewing distances from your monitor, you can see and read WORD clearly, but your vision rapidly descends into blurriness for anything not in your central cone of vision, so you have a hard time reading text above and below "WORD."

This is the reason why I wish that Eyefinity/NVSurround gave less processing power to the side monitors. I wouldn't mind if the central monitor got 90% of the processing power and the side monitors were blurry, because they are going to look blurry anyway when your vision is focused on the middle screen. Those side monitors there are just for immersion and for seeing MOVEMENT to the sides alerting you to possible important stuff, so that you can TURN YOUR HEAD to look at it. The human eye is tuned for detecting MOVEMENT very well, so well that we can see stuff moving even if it's far from the central cone of vision.

Now, if tessellation makes work easier for devs so they are more productive, I can see why that's a good thing, but in terms of visuals you can simply throw more polygons at the problem and get good results. (Obviously this only works up to a point.)

Don't get me wrong, I fully support tessellation and don't want us to go without it. But AMD's tessellation goes up to what, about 11x before it starts to fall behind NVs? I think 11x tessellation is probably fine. More would be better, if only very marginally so, but we live in a multiplatform world, so I doubt we'll see more than 16x tessellation on console ports unless NV pays the dev to include it. I'd be perfectly fine if we had up to 16x tessellation on all games from here on out. I doubt most people would see enough difference going from 16x to 64x to justify the performance hit, in most cases.

P.S. Lower on the page, NVIDIA folks brag about PhysX in the bank vault etc. but as gamers have discovered, turning PhysX on can decrease your enjoyment of the actual fights in games where PhysX is tossing lots of particles in the air and making it harder to see and fight as well as hurting framerates.

 

[Lonbjerg]

P.S. Lower on the page, NVIDIA folks brag about PhysX in the bank vault etc. but as gamers have discovered, turning PhysX on can decrease your enjoyment of the actual fights in games where PhysX is tossing lots of particles in the air and making it harder to see and fight as well as hurting framerates.

We could always go back to pong.
Added features cost performance.

I know some people whine over HDR...because it makes it harder to see.

I guess you can divide people into two groups.

1. People that want realism
2. People that really, deep down, want to play pong.

 

[sushiwarrior]

What you cant see gets culled in the Hull-Shader.

But higher tessellation factors waste rasterizer efficiency as well. Tiny polygons (high tessellation) only use a fraction of the rasterizers ability, yet take the same time to complete an operation (I assume).

 

[MrK6]

We could always go back to pong.
Added features cost performance.

I know some people whine over HDR...because it makes it harder to see.

I guess you can divide people into two groups.

1. People that want realism
2. People that really, deep down, want to play pong.

So you fall into the second group? Because PhysX is about as far from realistic as it gets. Unless nvidia lives in another dimension where an object can create 5x its bulk in debris.

 

[Imouto]

Tessellation was sold as a method to improve LOD capabilities in games making it more progressive. It turned out that it wasn't as effective as developers thought with a huge rule of diminishing returns in designing time, performance and visuals.

A combination of lower poly model + lower res and detail textures for LOD settings is still the preferred method and you can keep the performance in check.

Anyway crazy tessellation performance isn't needed except for absolutely dumb settings like Crysis 2 where the developer can screw up big time.

As for me tessellation isn't that important since it's not needed 95% of the time. You'd need to be badly focused on an object and really close to tell the difference between a high poly count model and a tessellated one. If you're focused on that model you don't need a high tessellation on anything else.

Those tessellation benches are pointless trying to measure impossible amounts of tessellation.

 

[blastingcap]

We could always go back to pong.
Added features cost performance.

I know some people whine over HDR...because it makes it harder to see.

I guess you can divide people into two groups.

1. People that want realism
2. People that really, deep down, want to play pong.

I like how you ignored 95% of my post to focus in on my postscript. I guess that means you accept the truth that you don't need insane amount of tessellation and that we're going to be console-limited anyway for most games.

 

[sontin]

But higher tessellation factors waste rasterizer efficiency as well. Tiny polygons (high tessellation) only use a fraction of the rasterizers ability, yet take the same time to complete an operation (I assume).

Nonsense.
Fermi+ chips have no problem working with 1x1 pixel size triangles.

Slide 23: http://www.highperformancegraphics.org/previous/www_2010/media/Hot3D/HPG2010_Hot3D_NVIDIA.pdf

 

[MrK6]

Nonsense.
Fermi+ chips have no problem working with 1x1 pixel size triangles.

Slide 23: http://www.highperformancegraphics.org/previous/www_2010/media/Hot3D/HPG2010_Hot3D_NVIDIA.pdf

Who cares? It's inefficient and impractical to render an image that way.

 

[sontin]

Seriously?!

Ah that must be the reason why we have so many great games which uses high poly models for terrain, surfaces and objectes.

But hey, marketing is marketing.

 

[Imouto]

Seriously?!

Ah that must be the reason why we have so many great games which uses high poly models for terrain, surfaces and objectes.

But hey, marketing is marketing.

It is a huge waste of resources and why we've got LOD for decades now. That's why you usually can find 3 different sets of meshes for each model. One for long distances, one for mid and a last one for close.

Tesselation was supposed to make the transitions between meshes for the same model less noticeable but hardware is powerful enough already to make mid and high poly ones pretty close. In fact with modern techniques you can make a low poly mesh out of a higher one really fast so there's no need to bother making complex tessellations and making them fit in a bigger scheme.

1px triangles are a waste anywhere and anytime.

 

[jj109]

Are you sugesting that you cannot tell non-tesselated from tesselated?

Honestly, I can't tell the difference from medium, high, and very high tessellation in games. There's only so many pixels and once the setting goes past medium, a lot of it is in the subpixel range.

 

[DiogoDX]

Honestly, I can't tell the difference from medium, high, and very high tessellation in games. There's only so many pixels and once the setting goes past medium, a lot of it is in the subpixel range.

When you reach some triangle number in a model adding more triangles make less impact in IQ.

 

[parvadomus]

Every arch has its pros and cons. Looking at the same techreport page, it looks like Hawaii has a good advantage in shader powa vs everything right now. I find it funny that full GK110 (5.3 tflops) can barely match Hawaii's 5.6tflops of shading powa in only 2 benchmarks.
I wonder what would happen if NV didnt pay developers to use insane amounts of tessellation. :awe:
Link

 

[Blue_Max]

I miss 3DFX Glide...

 

[BallaTheFeared]

Every arch has its pros and cons. Looking at the same techreport page, it looks like Hawaii has a good advantage in shader powa vs everything right now. I find it funny that full GK110 (5.3 tflops) can barely match Hawaii's 5.6tflops of shading powa in only 2 benchmarks.
I wonder what would happen if NV didnt pay developers to use insane amounts of tessellation. :awe:
Link

I don't know what you're talking about with the top part, but thanks for the link at the end.

I was having a really hard time trying to figure out what exactly stock 780 performance was and I think using ShaderToyMark might be a good way to figure that out.

It seems at 900/1500 I score 502 points which means 900MHz fixed is faster than whatever the 780 they used was.

 

[parvadomus]

I don't know what you're talking about with the top part, but thanks for the link at the end.

I was having a really hard time trying to figure out what exactly stock 780 performance was and I think using ShaderToyMark might be a good way to figure that out.

It seems at 900/1500 I score 502 points which means 900MHz fixed is faster than whatever the 780 they used was.

Then the card is throttling a little @ that benchmark. I wonder what a 290X fixed @ 1Ghz might do, maybe 700, 800 pts :|

 

[jj109]

When you reach some triangle number in a model adding more triangles make less impact in IQ.

The difference between high and very high tessellation in some games seems to be 60K triangles vs 600K triangles for all the good it does.

 

[jj109]

Then the card is throttling a little @ that benchmark. I wonder what a 290X fixed @ 1Ghz might do, maybe 700, 800 pts :|

Well, I just ran the benchmark and got 703 points on my overclocked GTX 780, which is where it should be compared to a stock 780 Ti. Power was only ~210 W according GPU-Z, so I doubt power or thermal throttling is involved.

Whoops this was supposed to be an edit

 

[parvadomus]

Well, I just ran the benchmark and got 703 points on my overclocked GTX 780, which is where it should be compared to a stock 780 Ti. Power was only ~210 W according GPU-Z, so I doubt power or thermal throttling is involved.

Whoops this was supposed to be an edit

I dont know, Im only trying to get a conclusion from his numbers.

 

[Vesku]

What does tessellation have to do with GPU vendor maintained DX 11 libraries?

Wonder if AMD will follow suit in addition to their Mantle API. Wouldn't be that hard to provide their Gaming Evolved partners with a few custom AMD controlled DX 11 libraries to go along with Mantle. They've already had to beef up their software team.

 

[ShintaiDK]

Wouldn't be that hard to provide their Gaming Evolved partners with a few custom AMD controlled DX 11 libraries to go along with Mantle. They've already had to beef up their software team.

You really dont think they dont already do that? Everyone supplies custom libraries.