Deus Ex Performance Thread

[Red Hawk]

In case anyone missed it, HardOCP posted their full performance analysis of the game. The AMD advantage from the preview is conspicuously missing; I wonder why? Nvidia holds the lead for the most part.

HardOCP came to the conclusion that tessellation actually isn't that much of a performance drain in this game, while SSAO is a major one. Same conclusion I drew in my (limited) benchmarking.

For what it's worth, the minor hit tessellation causes is less on AMD systems than Nvidia. A 580 had a 3% drop compared to a 2% drop on the 6970. So still, the idea that Nvidia is always better than AMD at tessellation is not necessarily true.

 

[AtenRa]

We can only have this kind of performance only if the game doesn't stress the hardware, witch means that DE3 doesn't implement a lot of tessellation and perhaps other DX-11 features.

No matter how efficient you make your code, performance will not be the same if you tessellate 5 objects and then 20.
Since we dont have any info to the contrary as of now, i believe DE3 only put a "mild" tessellation on humanoid models only, making it easier for the hardware.

http://forums.anandtech.com/showpost.php?p=32206782&postcount=264

I believe the same happening in most of the current DX-11 games

 

[Red Hawk]

The point remains though: what performance hit does happen is minorly less of a hit on AMD hardware than Nvidia hardware, rather than the opposite you'd expect if Nvidia had a definite tessellation advantage over AMD.

 

[AtenRa]

We do know that GTX5xx has a bigger performance hit in higher than 1080p res than HD69xx especially in 2560x1600, and since [H] once again, they only show us that high ress, it is possible that GTX580 lost more because of the higher ress and not of the Tessellation.

 

[Red Hawk]

No. The resolution isn't a variable in the tessellation comparison, the only thing that can be causing the difference is the tessellation or lack thereof.

 

[AtenRa]

Since we dont have data for 1920x1080/1200, we cant say why GTX580 lost more with Tess at 2560x1600.

 

[Red Hawk]

What other reason could there be? If Nvidia was worse off at that high of a resolution it should have been reflected in the no tessellation benchmarks already, and tessellation would have made no difference related to resolution. Anyways, the point is not that AMD is better than Nvidia with tessellation for the game at every resolution. The point is that it proves Nvidia does not have an unquestionable advantage in tessellation, since if it did then the performance hit would not be greater on Nvidia than AMD regardless of resolution.

 

[amenx]

The point is that it proves Nvidia does not have an unquestionable advantage in tessellation, since if it did then the performance hit would not be greater on Nvidia than AMD regardless of resolution.

There are varying degrees of tessellation. Heavy tess favors Nvidia, light tess (and other factors) may show different results.

 

[AtenRa]

Tessellating puts more work in other parts of the GPU as well, since Tessellation is not at the end of the Graphics pipeline, then perhaps GTX580 loses more because of rasters or texture units at high resolutions (2560x1600).

That is why i say we have to see if GTX580 loosing more than the HD6970 in lower resolutions (1080p etc) with tessellation or not.

If at 1920x1080 GTX580 loosing more than HD6970 again, then at that particular game it would seam that HD6970 tessellation units perform better, but if GTX580 loosing less than HD6970 at 1920x1080 then something else than Tessellation makes GTX580 loosing more at 2560x1600 resolution.

 

[StrangerGuy]

My 5850 runs this game fine enough, but I'm not convinced the visuals are technically worthy to justify a lot more GPU power required compared to like L4D2.

 

[3DVagabond]

That small of an amount difference could be Margin of error as much as anything else, don't you think?

 

[SirPauly]

That is good news.

Just a little follow up, since the game only support HD3D, others who use other means of 3D are left out. Of those, 3d vision users cried out loud in comparison. Usually Nvidia is quick on fixing things like that, but lately it seems to have dropping balls on game support. The one sticky post from Nvidia forum is lead by a mod asking the game dev to add 3d vision support nicely, which is, to me, finger pointing. I lost my hope on that game.

Until one day, a poster, rajkosto at Dx:HR forum, rajkoderp at Nvidia forum, came up with an interesting code, a dx code wrapper which hijacks the dx api calls and rerouted in a way that it works for 3D vision. In simple terms, HD3D puts the image to the left eye and the top and the right eye at the bottom and then render, where 3D vision requires the left eye image at the left, right eye image at the right + 1. In other word(Rajkosto's word), 2xheight for AMD, 2xwidth+1 for Nvidia. That is, if the calls can be hijacked, then the cost of re-ordering the left/right image is really small.

Well, it isn't surprising why rajkosto became the new hero on DX:HR, someone who does something for everyone for free. Although the code is not perfect, the fact it works lead to new questions.
A)If a user can write a code to fix it, then why Nvidia said they can't?
B)Why didn't the game add such a simple change before release?
C)Minimium depth, 0 convergence. Why?
D)"If the SLI engine was smart, it would figure out that the 2 renders to the 1x res rendertarget are independent, and do them in parallel, but unfortunately it isn't", rajkosto, Nvidia forum post 237.

Now he action is far more effective then forum partition post as he showed the possibility of making 3D work without having access to game code or driver, just redirecting calls that are suppose to end up at AMD's HD3D API into Nvidia's 3d vision with a little translation. That is:
Eidos: -1
Nvidia: -1
AMD : I am innocient.
rajkosto: +1000

With the latest patch, depth increases as the seperation increases and the announcement of 3d vision support is made.

Here is the site to get the wrapper, as well as detail usage and bugs with the wrapper.
http://rajko.info/hd3d/

I am hoping that the developer relations go much deeper than just getting Stereo3d to work but work well and solve some of the limitations with the help of nVidia and the developer. But yes, it is wonderful to see the community help improve things for others -- no doubt!

 

[Arkadrel]

The point remains though: what performance hit does happen is minorly less of a hit on AMD hardware than Nvidia hardware, rather than the opposite you'd expect if Nvidia had a definite tessellation advantage over AMD.

A 580 had a 3% drop compared to a 2% drop on the 6970.

We re talking about less than 1% differnce.

You see the Crysis 2 tesselation? nvidia drops like 20% and amd drops like 40% or something silly.

Are you saying AMD have influenced tessellation implimentation in a bad way in this game? and are doing it to make nvidia look bad? reeeeally? (to me it looks like nvidia does things like that, and not amd)

you'd expect if Nvidia had a definite tessellation advantage over AMD.

Why? because nvidia made you believe so? nvidia only has faster tessellation when it does absurb amounts of it. With small use of it, both are on equal grounds more or less.

Nvidia marketing works though, they just turn up tessellation to crazy levels, and go "look ours is faster".

Which is true... under insane levels (part most people forget).
(with no crazy levels of tess, things are much more equal)

you thinking anything tesselation related, nvidia has to be faster, is because nvidia marketing got you thinking that.

 

[SirPauly]

It would of been nice to see more tessellation, like landscapes but I wonder why it wasn't added?

 

[RussianSensation]

The AMD advantage from the preview is conspicuously missing; I wonder why?

This is explained clearly in the article.

"The version of the game we received for our preview article was a reviewer build [reason 1], which was an earlier build than the launch version. On top of that, the game has been updated twice since launch [reason 2], with two patches, one on day one and one on day two of release. Also, on top of those facts our level we tested on is different [reason 3] along with the intensity and complexity of game elements tested are much more greater now [reason 4], than they were in the preview."

Nvidia holds the lead for the most part.

44.3 vs. 39.9 fps (WOW, 10% faster for 30% higher cost)

"GTX 580 was consistently about 10% faster than the Radeon HD 6970. Of course, this doesn't quite balance with the fact that the GTX 580 is currently about 30% more expensive than the Radeon HD 6970."

But wait, there is more, for 6+ months now, you could have purchased an HD6950 and unlocked it in about 5 minutes into an HD6970 with a 95%+ success rate. Those cards generally went for $230-250. I think most people here would choose a 10% performance loss and $200 in their pocket towards their next GPU upgrade.

 

[SirPauly]

I found my answer:

GameSpy: Any features you wanted to work into the PC version that didn't make the cut? And of those features that did make it, which do you think will impress PC gamers the most?

Jurjen Katsman: As for something that didn't make the cut, after we had been quite successful in using tessellation on characters, we also spent quite some time trying to use it on the environment as well. With the type of cityscape environments that we see in Deus Ex: Human Revolution it ultimately wasn't a good fit.

http://pc.gamespy.com/pc/deus-ex-3-unofficial-title/1187766p1.html

 

[Makaveli]

This is explained clearly in the article.

"The version of the game we received for our preview article was a reviewer build [reason 1], which was an earlier build than the launch version. On top of that, the game has been updated twice since launch [reason 2], with two patches, one on day one and one on day two of release. Also, on top of those facts our level we tested on is different [reason 3] along with the intensity and complexity of game elements tested are much more greater now [reason 4], than they were in the preview."



44.3 vs. 39.9 fps (WOW, 10% faster for 30% higher cost)

"GTX 580 was consistently about 10% faster than the Radeon HD 6970. Of course, this doesn't quite balance with the fact that the GTX 580 is currently about 30% more expensive than the Radeon HD 6970."

But wait, there is more, for 6+ months now, you could have purchased an HD6950 and unlocked it in about 5 minutes into an HD6970 with a 95%+ success rate. Those cards generally went for $230-250. I think most people here would choose a 10% performance loss and $200 in their pocket towards their next GPU upgrade.

^^^^ Bingo!

I am one of those 6950---6970 users!


/thread

 

[Red Hawk]

That small of an amount difference could be Margin of error as much as anything else, don't you think?

That is a fair point; my point is more in the absence of an Nvidia advantage rather than the existence of an AMD advantage.

A 580 had a 3% drop compared to a 2% drop on the 6970.

We re talking about less than 1% differnce.

You see the Crysis 2 tesselation? nvidia drops like 20% and amd drops like 40% or something silly.

Are you saying AMD have influenced tessellation implimentation in a bad way in this game? and are doing it to make nvidia look bad? reeeeally? (to me it looks like nvidia does things like that, and not amd)

No, I'm not...

Why? because nvidia made you believe so? nvidia only has faster tessellation when it does absurb amounts of it. With small use of it, both are on equal grounds more or less.

Nvidia marketing works though, they just turn up tessellation to crazy levels, and go "look ours is faster".

Which is true... under insane levels (part most people forget).
(with no crazy levels of tess, things are much more equal)

you thinking anything tesselation related, nvidia has to be faster, is because nvidia marketing got you thinking that.

Again, no -- where I heard the idea that Nvidia has better tessellation performance is right here, on this board, from certain members. Certain members who are not 100% right, as this case shows.

 

[SirPauly]

nVidia's strength is based on the higher the maximum level -- the more the two sides separate. With lower settings, AMD does fine.

 

[Red Hawk]

HardOCP did a performance and image quality analysis of Human Revolution's latest DLC, The Missing Link. Apparently, the DLC improved the tessellation models and the lighting effects significantly over the main game. Unfortunately, the improved tessellation detail results in AMD cards chugging somewhat on tessellation heavy scenes.

 

[Dribble]

But the better graphics aren't in the main game?

Surely they should port them back and patch the real game everyone wants to play - or have ati paid too much and hence they are not allowed to add more tessellation that makes ati cards look bad.

 

[Silverforce11]

HardOCP did a performance and image quality analysis of Human Revolution's latest DLC, The Missing Link. Apparently, the DLC improved the tessellation models and the lighting effects significantly over the main game. Unfortunately, the improved tessellation detail results in AMD cards chugging somewhat on tessellation heavy scenes.

Saw that, good review. In cutscenes when tessellation is to the max, NV wins by a big margin.