AMD: R8xx Speculation

[Silent_Buddha]

Probably, although the meshes don't seem to be that dense in the wireframe parts on the youtube video. As long as a triangle is larger then 4-6 pixels, it should not cause that much trouble.

Some thorough testing could answer the question, run the demo with tesselation, but in a relatively low resolution, run it with and without SSAO, and so on... I bet there'll be a tech website with an article soon.

If I get time maybe I'll take a screenshot and upload it.

However, as long as you don't look at the Dragon Statue or sides of buildings, triangle sizes are generally greater than 4-6 pixels.

But, the stones in the corner of buildings and most especially the Dragon Statue. Triangles do go to sub-pixel sizes. Put into wireframe and look at the Dragon statue and most of it is just a solid mass of white. Even when viewed up close, large portions of the dragon are solid white.

That's why I said, the use of tessellation on the statue is excessive and probably meant primarily to stress the engine/video card.

Doors also get an excessive level of tessellation that is totally unwarranted.

There's a LOT of optimisation that could be done with regards to excessive tessellation on objects that don't need as much.

The cobblestone pathway while exaggerated uses an order of magnitude less triangles than the flat doors for instance. Reduce the number of triangles on the flat doors by an order of magnitude and performance would increase greatly.

Regards,
SB

 

[Laa-Yosh]

The problem is that displacement mapping generally only looks good with subpixel triangle sizes. Otherwise the results get... noisy.

 

[Silent_Buddha]

Some screens to illustrate...

Door and Cobblestone road.



Dragon Statue...




A bit excessive on both. And kills performance.

The cobblestone road and roofing shingles on the other hand show tesselation extremely well with very few triangles comparatively.

But again things like that would be up to the software dev to decide when they use this engine to make their game.

[edit]Swapped out image host, hopefully it worked right.

Regards,
SB

 

[fellix]

Adaptive tessellation is nice -- one step beyond a brute LOD scaling would be redistributing the subdivision factors between front facing surfaces and those below a certain steep angle. The last one definitely needs more tessellation to keep the detailed silhouette definitive than the former, if you understand what I mean.

 

[Kaotik]

Some screens to illustrate...

Small warning, one of apparently random imageshack adverts launches "fake virus scan" advert, which is infact trojan downloader, MSE instantly popped up, it's TrojanDownloader:JS/Renos with alert level "high"
This is as far as I know completely random advert so just a warning if anyone of you happen to hit it

edit:
I suggest using any other imagehosting service instead of that; heck, if you mail me the images I can host them for you
one free that I know to be virus-free is aijaa.com, though it's in finnish which can cause problems to some - the controls should be rather obvious for everyone though

 

[Mat3]

To me, the real benefits of tessellation won't show up until DX11 is a minimum requirement of a game. The tessellation being added to current games seems to be mostly taking a medium-high detail model and making it super detailed. Looks nicer up close, but overall won't doesn't make a big difference while playing.
But if designed from the ground up, and not worrying about hardware that doesn't support it, almost everything can be a low poly model made into a high detail one on the fly, and with that, major memory savings. That should translate into larger game environments while at the same time having more detailed objects and characters. Right now, games typically sacrifice area size or detail, or re-use same models of trees and characters throughout ad nauseam. Tessellation can make all that go away.

 

[Laa-Yosh]

As I've tried to indicate already - tesselation doesn't necessarily mean that you can use lower polygon counts. In fact, you usually have to use more polygons to build proper models.

 

[Silent_Buddha]

Small warning, one of apparently random imageshack adverts launches "fake virus scan" advert, which is infact trojan downloader, MSE instantly popped up, it's TrojanDownloader:JS/Renos with alert level "high"
This is as far as I know completely random advert so just a warning if anyone of you happen to hit it

edit:
I suggest using any other imagehosting service instead of that; heck, if you mail me the images I can host them for you
one free that I know to be virus-free is aijaa.com, though it's in finnish which can cause problems to some - the controls should be rather obvious for everyone though

If you want you can just download the images from imageshack and host them elsewhere, I can redo the links after.

Regards,
SB

 

[Psycho]

Q6600 @ 3GHz, ATI 5850 stock
1600x1200 0xAA 4xAF DX11 Tesselation 33.5 fps

E6750 @ 3.2GHz, HD 5770, 1600x1200 0xAA 4xAF DX11 Tesselation.
850/1200mhz (stock): 24.8 fps
850/1400mhz: 25.4 fps
950/1400mhz: 27.8 fps

Not much bandwidth limit here

Btw, is it only me that get a "win7 required" message on install of the ati demos? (on my win7 rc)

 

[sc3252]

Someone with 5800 card to cap FPS numbers with FRAPS during the timedemo (tessellation on/off), and plot them as nice graph in Excel, please!
Or just gimme the FRAPS output dump and I'll do the Excel thing.

This is a crappy line graph of tessellation on and off. its run at 1680x1050 af at 8 and every thing at max except aa which is off. I ran it twice and took the second numbers of each setting.
Note: this is not a stock clock 5850 its @775/1100

hopefully this works, I havent really used photo hosting sites before.
If someone wants to see the fps data in detail.
tessellation on

60
59
57
55
52
52
50
49
48
49
50
59
59
58
56
56
54
49
50
48
46
48
49
50
50
49
51
50
49
52
52
52
52
45
36
37
41
42
41
35
33
32
33
33
33
65
65
70
69
69
51
35
32
32
33
35
40
48
46
46
41
44
63
64
65
65
65
50
35
34
34
31
27
25
21
21
18
17
18
24
26
24
22
21
20
21
19
19
21
23
27
27
27
27
27
27
23
21
19
21
21
23
27
30
30
28
26
25
23
21
22
22
25
19
16
16
18
19
18
18
17
17
17
16
19
25
25
27
27
26
27
24
22
21
20
30
38
41
43
48
51
51
47
46
37
33
36
40
40
39
35
33
29
25
26
28
30
26
19
19
17
17
18
21
22
28
28
28
28
26
24
24
24
23
25
27
29
26
23
22
28
29
30
33
29
27
25
24
22
22
21
21
19
20
19
19
17
21
28
28
28
27
27
29
28
28
28
28
27
26
25
26
26
27
28
30
31
33
33
37
69
71
69
65
61
52
45
64
59
48
36
42
47
64
64
64
65
63
45
44
45
46
37
32
31
30
30
32
27
25
23
23
25
32
57

Tesselation off

69
66
64
61
59
55
51
56
52
53
53
61
61
60
58
59
55
50
51
50
47
50
50
51
52
51
52
53
51
52
54
54
54
49
54
54
56
58
59
44
41
43
42
42
39
72
66
72
71
71
61
53
55
54
52
50
48
50
49
48
44
46
69
71
70
71
72
64
57
60
62
62
65
64
65
64
63
65
65
53
52
53
55
56
59
61
57
54
50
50
76
85
85
86
88
78
64
60
59
62
63
64
64
65
64
64
64
64
64
65
65
65
62
59
62
62
60
63
60
58
59
60
59
60
57
62
64
62
60
56
60
59
59
59
57
47
46
49
52
56
57
57
57
58
53
54
55
56
54
52
50
46
44
43
40
45
47
48
64
60
61
61
60
61
61
62
62
63
63
64
66
66
67
67
69
70
74
72
70
68
58
69
70
74
74
73
72
68
70
68
65
63
64
64
63
68
61
54
51
51
49
48
48
48
47
46
47
47
47
60
60
60
60
59
60
61
61
60
60
60
78
84
83
80
75
59
68
70
69
67
68
48
60
74
77
79
79
75
53
56
56
57
60
73
72
71
69
72
73
74
74
68
69

 

[Kaotik]

If you want you can just download the images from imageshack and host them elsewhere, I can redo the links after.

Regards,
SB

http://home.akku.tv/~akku38901/B3D/hosted/

There's pics including those thumbnails if you want to use them

 

[Silent_Buddha]

K...I'll swap them out. No clue how to do it, hopefully just swap in the urls for the ones in my post? Will give it a shot.

Regards,
SB

 

[Silent_Buddha]

As I've tried to indicate already - tesselation doesn't necessarily mean that you can use lower polygon counts. In fact, you usually have to use more polygons to build proper models.

I don't doubt that, I was just pointing out that for this demonstration at least, it seems there are some objects with excessive tessellation even though I'm not sure it's needed. The door for example, if you look at it even up close you can't tell it's anything other than a flat texture until you turn on wireframe. And that causes some rather extreme slow down.

While other area's show very effectively the potential of tessellation even without an excessive level of tessellation. Even if the effect itself is a bit exaggerated (cobblestones).

In the end it'll be up to whatever dev uses this engine to determine how much to use.

Regards,
SB

 

[Laa-Yosh]

I'm talking about the pre tesselation meshes here. If every triangle gets smoothed, you loose tight corners and sharp edges - unless you add smaller strips of polygons there to maintain the shape. But that increases the base polygon count significantly (and is also the reason why cars in racing games are better off without it).

 

[Mintmaster]

Probably, although the meshes don't seem to be that dense in the wireframe parts on the youtube video. As long as a triangle is larger then 4-6 pixels, it should not cause that much trouble.

4-6 pixel triangles will usually cover four quads, so that's far from efficient. Even if you have favourable shapes (i.e. not skinny), you'll need more than 100 pixels per triangle to break 80% utilization in the shader/pixel pipes.

Moreover, tesselation increases the time spent limited by setup, depending on how complex the pixel shaders are.

 

[gongo]

I have a strange bug with 4870 1GB and Dell 2407 monitor.
I select DX10 1920x1200 (16:10) but the demo only outputs to 1600x1200 @ 59hz (4:3)? It happened to me with 3d mark Vantage too...could never test at 1920x1200...

Am i running out of memory?

 

[neliz]

Am i running out of memory?

what has your memory size to do with output resolution?

 

[Lightman]

I've stumbled across this interesting website:
http://users.skynet.be/fquake/

Really nice DX Compute Demos!

From description off one of them:

Here a quite fast Mandelbrot and Julia viewer, making use of DX11 and the DirectCompute API.



The set is calculated with up to 1024 iterations. Making use of the horsepower of DX11 GPUs enables real-time panning and zooming even at high resolution.



Two versions are included, a scalar one and a vectorized computation version.

Both generate the same output. The vectorized version was made after suboptimal performance on the ATI HD 5870 with scalar calculation.

Compared to the scalar version it runs twice faster on this GPU. Here we see the backside of a non scalar GPU architecture. On forthcoming scalar Nvidia DX11 GPUs, probably the scalar version will run faster. Writing a vectorized compute shader is substantial more complicated.

On the HD 5870 computational throughput is over 1.9 TFLOP/s.



Full source code is included.

Remark that no drawing code was needed. It is possible to directly write to the backbuffer from the compute shader.

 

[Laa-Yosh]

4-6 pixel triangles will usually cover four quads, so that's far from efficient. Even if you have favourable shapes (i.e. not skinny), you'll need more than 100 pixels per triangle to break 80% utilization in the shader/pixel pipes.

Moreover, tesselation increases the time spent limited by setup, depending on how complex the pixel shaders are.

Yeah, it seems that's the catch with the DX11 approach and the current hardware.

Good looking displacement actually needs subpixel sized triangles... and there we enter a whole new world of aliasing and sampling issues. I guess there's a reason why Tim Sweeney's talking about micropolygon rendering and software based engines - they've probably did their own displacement tests with UE4 and Gears assets already. So they know very well that this is going to be a problem.

 

[OpenGL guy]

I've stumbled across this interesting website:
http://users.skynet.be/fquake/

Really nice DX Compute Demos!

See http://forum.beyond3d.com/showthread.php?t=55330