GTX680 vs HD7970 with SGSSAA, user input requested

[boxleitnerb]

Good morning guys,

I've been a bit baffled by some results I've seen when using SGSSAA. Basically it seems that the 7970 is unusually weak with SGSSAA. I thought we could investigate this by benchmarking some games. I gathered some results from other places already - the driver versions might not be up to date always, but the results are still quite unusual I think.

Dirt Showdown, 2560x1440 all max, 4xSGSSAA, 16xAF

GTX680@stock, FW310.61

HD7970@1125, Cat 12.11b3

Hitman Absolution, 2560x1440, 8xSGSSAA, 16xAF

GTX680@stock, FW310.61

HD7970@1200, Cat 12.11b8

Hitman Absolution, 1920x1080, 8xSGSSAA, 16xAF

GTX690@1150, FW310.61

HD7970@1150, Cat12.11b8

Alien vs Predator 3, 1680x1050 (Radeon) and 1920x1080 (Geforce), 4xSGSSAA,16xAF

GTX690 +150, FW306.02 HQ

HD7970@1330, Cat 12.8

Edit:
Is it possible to place two screenshots beside each other, that wouldn't waste that much space.

 

[aaksheytalwar]

These bechmarks mean nothing. You need to be borderline playable to have benchmarks which makes sense. If one card gets 5 fps avg and the other card 7 fps avg then the second one is not 40% faster.

Your results are frivolous at best.

You need to make sure that one card gets at least 50ish or 60ish fps to compre the other card.

 

[boxleitnerb]

These were the only results I could find with SGSSAA - this is why I have asked for others to join in to gather more values
Also consider that with SLI/CF the fps double or triple. While I agree that around 60fps avg are necessary for most gamers, that doesn't invalidate the results. I hope for Hitman I'll be able to post results at 1080p later today which should increase fps considerably.

Would you consider running some benchmarks for this thread, please?

 

[Granseth]

I think you are right in saying that 7970 delivers poor results using SGSSAA. Thats my experience too, but I only have a 7970, so I can't compare to a 680.

 

[boxleitnerb]

That's not a problem, we could compare with the 680 from someone else as long as drivers are up to date. Just post some benchmarks if you want, preferably playable settings and games with ingame support for MSAA so we don't have to force AA.

 

[Granseth]

OK, will see what I can do. But the only two games I can remember now that have ingame support is Portal 2 and Skyrim, but my Skyrim game is heavily modded so thats not comparable. And I think you can get Portal 2 results from Anandtech's reviews.
But I probably have other games with ingame support, I just have to find them first

 

[boxleitnerb]

There are some standalone benchmark tools like for Alien vs Predator (DX11) or Far Cry 2 (DX10) or Call of Juarez (DX10).

Metro 2033, Shogun 2, BF3, Alan Wake, Just Cause 2, Max Payne 3 etc. all have ingame MSAA. Not all of them have integrated benchmarks, though.

 

[Granseth]

Shogun 2 and Just Cause 2 is at my list of games at least Will try to run a bench tonight or tomorrow.

 

[Jaydip]

I can do some testing using AC3 if you want.

 

[boxleitnerb]

Shogun 2 and Just Cause 2 is at my list of games at least Will try to run a bench tonight or tomorrow.

Great, thank you. Both games have integrated benchmarks. Shogun 2 has one, Just Cause 2 has three. Please remember to post the settings as well. Maybe two runs each, 4xMSAA and 4xSGSSAA so we can see the performance cost.

I can do some testing using AC3 if you want.

Okay, thanks

 

[lehtv]

Help a noob out: how do I enable SGSSAA? Is it just the Super-sample AA option in CCC? I could do some benches on my OC'd 7950 using 12.11b8

 

[boxleitnerb]

The AA-quality slider has three settings:

Left: MSAA
Middle: MSAA+AAA (transparency)
Right: SGSSAA

Nvidia users just select it via NV Inspector in the "Antialiasing - transparency supersampling" entry:

On AMD cards, the SGSSAA samples always match the MSAA samples automatically. So if you select 4xMSAA ingame, you get 4xSGSSAA. On Nvidia cards, you are a bit more flexible and can select the same amount of samples or lower. So 8xMSAA would allow for 8x, 4x and 2x SGSSAA, for example. But for the benchmarks in this thread, please select matching samples, too.

 

[lehtv]

Sapphire 7950 3GB @ 1115 / 1330
Catalyst 12.11 beta 8
1920x1080

AvP Benchmark 1.03

Default settings (no AA)
Average: 100.4

4x MSAA
Average: 70.0

4x SSAA
Average: 23.9

Heaven 3.0

Tessellation: Normal
AF: x16
AA: x4

4x MSAA
Min: 28.0
Avg: 71.5
Max: 151.0

4x SSAA
Min: 19.0
Avg: 37.9
Max: 87.8

 

[boxleitnerb]

del

 

[boxleitnerb]

GTX580@stock, FW310.61
1080p

4xMSAA: 48.7
4xSGSSAA: 17.9

Interesting. With MSAA, the 7950@OC is 43% faster, but with SGSSAA, it is only 33% faster.

 

[lehtv]

Added Heaven 3.0

 

[DiogoDX]

SGSSA is diferent of SSAA. Radeons don't do SGSSAA but traditional SSAA wich is heavier.

 

[lehtv]

Is SSAA better quality than SGSSAA?

 

[boxleitnerb]

SGSSA is diferent of SSAA. Radeons don't do SGSSAA but traditional SSAA wich is heavier.

Nope, Radeons do SGSSAA. SGSSAA is a subset of SSAA, OGSSAA is also a subset of SSAA. SSAA is just a general term and does not mean a specific method.

2x2 OGSSAA (ordered grid = not rotated, not sparsed):

4x SGSSAA (sparse grid):

Is SSAA better quality than SGSSAA?

Depends on different factors. Usually SGSSAA looks significantly better. The EER (edge equivalent resolution) of 4xSGSSAA can only be reached by using 16xOGSSAA (4x the resolution of each axis). That would be as demanding as running a resolution of 7680*4320. 1920x1080 + 4xSGSSAA is certainly much faster while looking equally good.
But when MSAA is badly implemented, leaving out polygon edges like in Hitman or Max Payne 3, OGSSAA via downsampling/downscaling can look better overall, because it works on every part of the image.

 

[DiogoDX]

Nope, Radeons do SGSSAA. SGSSAA is a subset of SSAA.

The story behind SGSSAA:
Nvidia implemented TRSSAA as an alpha test that supersampled pixels that passed. A bug was present in one of the drivers they released for the new fermi cards that caused the alpha test to always pass and therefore apply SGSSAA to all pixels. They fixed the bug in the next release but many users said that they liked the look of fullscene SGSSAA and complained about its removal. This caused nvidia to release a tool that allowed the bug to be re-enabled. Despite the fact that this tool was made by nvidia they do not officially support it, it comes with a nice fat "use at your own risk" warning. The new fermi cards + new drivers made nhancer no longer compatible (since it never got updated for the 200 series drivers) and since the developer of nhancer has dropped off the face of the planet we can expect it to never be updated. However in the meantime another developer made a better app that did the same thing, nvidia inspector. He included the SGSSAA hack in it.

Therefore SGSSAA technically comes in two forms, FSSGSSAA (fullscene sparse gride supersampling anti-aliasing, which is SGSSAA applied to the entire scene) or TRSGSSAA (tranparency sparse grid supersampling anti-aliasing, which is SGSSAA applied only to transparent textures).
As far as nvidia inspector and nvcp (nvidia control panel) are concerned they are listed under the following names:
FSSGSSAA is just called SGSSAA
TRSGSSAA is just called TRSSAA or "Transparency Supersampling"

Technically both of them are SGSSAA, but only the fullscene implementation is called SGSSAA.

http://www.overclock.net/t/1250100/nvidia-sparse-grid-supersampling

 

[DiogoDX]

Nope, Radeons do SGSSAA. SGSSAA is a subset of SSAA, OGSSAA is also a subset of SSAA. SSAA is just a general term and does not mean a specific method.

2x2 OGSSAA (ordered grid = not rotated, not sparsed):

4x SGSSAA (sparse grid):

Depends on different factors. Usually SGSSAA looks significantly better. The EER (edge equivalent resolution) of 4xSGSSAA can only be reached by using 16xOGSSAA (4x the resolution of each axis). That would be as demanding as running a resolution of 7680*4320. 1920x1080 + 4xSGSSAA is certainly much faster while looking equally good.
But when MSAA is badly implemented, leaving out polygon edges like in Hitman or Max Payne 3, OGSSAA via downsampling/downscaling can look better overall, because it works on every part of the image.

This pictures shows OSGSSAA and RSGSAA not SGSSAA.

OGSSAA = SSAA = downsampling entire image
RGSSAA = Rotated Grid Super-sampling Anti-aliasing:

Similar to SSAA. The difference is that instead of the sample grid being along horizontal and vertical axis, the grid is rotated to a fixed angle. This causes better anti-aliasing of edges that are almost horizontal or vertical.

SGSSA = Sparse Grid Super-sampling Anti-aliasing

Similar to SSAA. This type of AA is the first in a series of AA types that represent a compromise designed to trade image quality for performance.

The sampling pattern of plain SSAA is called Regular Grid or Ordered Grid. It can be pictured as a grid that has higher resolution than the pixel grid. Where SSAA samples in each square in that grid, SGSSAA samples in only some of the squares.

http://www.dahlsys.com/misc/antialias/index.html

 

[boxleitnerb]

First: There is no "SSAA". This term is incomplete. The first two characters, that denote the sample arrangement, are missing. It has to be either OGSSAA, RGSSAA or SGSSAA.
You said, Radeons don't do SGSSAA, but "traditional SSAA, which is heavier". That is not correct. They certainly don't do OGSSAA. Most of the time, OGSSAA does not trade image quality for performance. OGSSAA is a brute force method and looks worse than SGSSAA or RGSSAA at comparable performance.

I selected SGSSAA on my 580 to obtain the second picture. If it is SG or RG depends on how you want to look at it. It can be seen as rotated, but also as sparsed.

GTX580@stock
Heaven 3.0

Tessellation: Normal
AF: x16
AA: x4

4x MSAA
Avg: 53.5

4x SGSSAA
Avg: 27.1

Here everything seems normal. Unless the card is CPU bottlenecked with MSAA.

This just came in from a buddy of mine:
GTX680@stock@310.61
AvP3, 1080p, HQ-AF, 4xSGSSAA

Average: 23.4

I think I should make a table for better clarity.

 

[DimmyK]

GTX 680 @ 1280/3335, 310.61

AvP benchmark, 1080p

4xMSAA - 65.5

4xSGSSAA - 27.2

 

[boxleitnerb]

Thx! Could you benchmark Heaven 3.0, too? Or what other games do you have?

Lets stick to 1080p for all future benchmarks, makes it easier for me to put it all together in a table. Thanks in advance.

 

[Jaydip]

Heaven 3.0 GTX 680 @ 1208/6418 310.54

Tessellation: Normal
AF: x16
AA: x4

4x MSAA
Min: 38.1
Avg: 80.6
Max: 183.2

4x SSAA
Min: 24.9
Avg: 42.5
Max: 105.6