ATI's DX11 tessellation to be aided by the shader core?

[MistaPi]

http://www.overclockers.com/forums/showpost.php?p=6477007&postcount=8

Is there any truth to this?

[Picao84]

Hmm.. Some sentences seem a bit naive to me.

Nvidia uses 16 tessellation units that are emulated within the CUDA cores.

This was what Charlie said. It was never proved that there isnt any dedicated hardware for Tesselation in GF100.

Currently, ATI is working on changing this within the drivers so that the Stream Processors will aid the Tessellation Unit with processing tessellation.

Hello? They are already aiding the Tesselation Unit by doing Domain and Hull Shaders.

They're also working to further improve tessellation by making the cache work better between the tessellation unit and the SP's, which is going to be in the 11.0 Catalyst drivers.

Isnt this correlated with hardware? If the hardware was intended to be like that, drivers would have this long ago. They would not be waiting for Fermi to come out. Maybe this will be addressed in SI, but i doubt something like this can be done in Evergreen just by software tweaks.

[MDolenc]

And I thought the silly season was over...

[willardjuice]

I heard this is coming out right after dx9.1

[Psycho]

which is going to be in the 11.0 Catalyst drivers.
:roll:

[ShaidarHaran]

:roll:

What, your calendar doesn't have Nulluary?

[digitalwanderer]

I heard this is coming out right after dx9.1
http://www.elitebastards.com/forum/images/smiles/biglol.gif

I get it! :lol:

[nbohr1more]

I loved that DX 9.1 rumor... It was so deliciously desperate... The guys who released and fixed the original Longhorn OS to show "what could've been" should make sure they include a DirectX version that is "optimized for the FX architecture" like 9.1 was supposed to be...

[Arnold Beckenbauer]

http://www.overclockers.com/forums/showpost.php?p=6477007&postcount=8

Is there any truth to this?

No, lies, lies, lies.

The truth is: Cypress has already four TUs, but 3 of them are disabled for lower TDP. But when tessellation will become a standard, all TUs will be enabled.

10.5 will rock!!!! (http://forums.guru3d.com/showthread.php?t=321784)

[psolord]

Just for the record, I gave it a shot with 10.5 beta on my 5850 crossfire and I saw no great difference. Just a couple of frames, perfectly within the error margin.

[Novum]

What was to be expected...

[Arnold Beckenbauer]

Just for the record, I gave it a shot with 10.5 beta on my 5850 crossfire and I saw no great difference. Just a couple of frames, perfectly within the error margin.

What 10.5 beta? :smile:
But not these betas: http://www.rage3d.com/board/showthread.php?t=33961625?

Seriously: Are there no ways to improve Evergreens' tessellation performance? Shader replcements, better LDS usage, etc?

[psolord]

No it was this release.
(http://www.ngohq.com/news/17660-ati-catalyst-10-5-beta.html)

[Grall]

Generally, 3D hardware vendors don't retrofit new features onto existing products... They want us to buy new hardware entirely instead, so that they'll make more money. :) Also, it's probably easier to just improve tesselation in the next generation of GPUs than to try and fix the current product line.

My wager is that as usual, wild internet rumors are just that. ;)

[Dave Baumann]

Generally, 3D hardware vendors don't retrofit new features onto existing products...

I would contend that, actually. Generally speaking, where we can, we try and backfill features to the prior gen where they are capable of doing it. Occasionally there have been some cases where the feature has been qualified first on a newer set of hardware an there has been a lag in getting those new features on the old stuff. An example of this is CCC recently added some new video capabilities and they are not Evergreen specific.

[Grall]

On the other hand, ATI announced a few months back you guys would not spend much effort on improving OpenCL performance for the 4xxx series. So the door definitely swings both ways. :razz:

[Dave Baumann]

And yet OpenCL as a feature has been retrofitted into RV7xx.

[Arnold Beckenbauer]

And yet OpenCL as a feature has been retrofitted into RV7xx.
Without image support.

The real ugly thing, that you did, was the removal of SGSSAA under OpenGL for HD4000 owners. This is how it should work: http://nvidia.custhelp.com/cgi-bin/nvidia.cfg/php/enduser/std_adp.php?p_faqid=2624

[Grall]

The real ugly thing, that you did, was the removal of SGSSAA under OpenGL for HD4000 owners.
Don't care about that, or anything SS-related at all really. What I REALLY miss is them removing alpha to texture type "antialiasing" for billboard/foliage type polygons and replaced it with supersampling that even on a crossfire 4890 setup is so damn slow it's totally unusable even in titles like WoW that's half a decade or more behind the curve on a technical level.

Alpha to coverage looked great, ran fast. I don't understand at all why the fuck it was removed and replaced with something that's just total crap that simply can't be used. Forget Crysis... The GPU would have a coronary infarction just thinking about it. :sad:

Other than the extreme idle power useage of the 4890, this is the only thing that really bugs me with those cards. Well, the crappy texture AF too and the less effective MSAA compared to G80 (particulary on lines, which don't seem to be AA'd at all), but that I can ignore more easily.

However, WoW on my now almost ancient G80s in SLI, 16x AA, alpha to coverage turned on, max AF...looks awesome, runs slick as snot. ATI 8x AA + max AF is noticeably more jagged, + textures shimmering where they do not on the G80s, and of course no alpha to coverage, so crawling ants around the edges of every damn tree and plant. That sucks.

[Davros]

A noob question if I may:
tessellation performance on ati 5000 series is poor yes ?

[Lightman]

A noob question if I may:
tessellation performance on ati 5000 series is poor yes ?

Compared to Fermi with high tesselation factor - yes.
Compared to XBOX360 - very fast

[Arnold Beckenbauer]

Alpha to coverage looked great, ran fast. I don't understand at all why the fuck it was removed and replaced with something that's just total crap that simply can't be used. Forget Crysis... The GPU would have a coronary infarction just thinking about it. :sad:
You mean EATM, don't you? EATM was really cool and damn fast.

Other than the extreme idle power useage of the 4890, this is the only thing that really bugs me with those cards. Well, the crappy texture AF too and the less effective MSAA compared to G80 (particulary on lines, which don't seem to be AA'd at all), but that I can ignore more easily.
The AF isn't crappy, it's optimised. :lol:



Compared to Fermi with high tesselation factor - yes.
Compared to XBOX360 - very fast
Actually it's slower: HD5000 needs three cycles to generate a tessellated triangle, Xenos two cycles.

[BRiT]

Actually it's slower: HD5000 needs three cycles to generate a tessellated triangle, Xenos two cycles.

But isn't the HD5000 clocked higher and has more units, so net effect is it's faster?

[Lightman]

But isn't the HD5000 clocked higher and has more units, so net effect is it's faster?

Yeah, no one said clock for clock :razz:

[3dcgi]

Actually it's slower: HD5000 needs three cycles to generate a tessellated triangle, Xenos two cycles.
If you implement Xenos style tessellation HD5000 and 4000 are actually slightly more efficient clock for clock due to some changes around the tessellator. This is part of why the HD4000 series claimed improved tessellation. The other part was GS support. DX11 tessellation is different.

[Grall]

You mean EATM, don't you?
I suppose I do, I don't know the exact term for it, but the actual result of the technique looked just like Humus' ancient alpha-to-coverage DX9 demo to my eyes. It might have had a slight performance hit I'm not sure, but it was basically nothing worth complaining about, particulary compared to the two supersampling offerings currently in the drivers - one of which barely gives any improvement at all at a big performance sock in the jaw, and the other a slight improvement at a horrific knock-out punch... Terrible! :(

The AF isn't crappy, it's optimised. :lol:
Well, it sure is something, that's for sure! :D

I'm very impressed with the 5000 series' AF tho, I wish I could warrant splurging for one of those cards, but not this late in the product's life cycle. I'll wait and see for what they have in store for the 6000 (guessing) series, betting my money on that they don't screw up the AF again, heh.

Hopefully there'll be a lot more tesselation oomph in those chips, and maybe they'll even hit the market as some software starts using tesselation in a meaningful way, who knows? :)

...Perhaps the 6870 will also finally be able to run Crysis at playable fps with antialias turned on, lal...

[DavidGraham]

http://alienbabeltech.com/main/?p=12648&page=2
http://alienbabeltech.com/main/?p=12648&page=3

[Davros]

I'll rephrase my question would the speed of tessellation be an issue in games on a 5000 series

[willardjuice]

I'll rephrase my question would the speed of tessellation be an issue in games on a 5000 series

Debatable. I suppose eventually it will catch up to the r8x0, but will that happen during its lifespan? At the moment, it doesn't appear so.

[Davros]

And am I right in thinking all 5000 series cards have the same number of tessellation units ?

[Dave Baumann]

Debatable. I suppose eventually it will catch up to the r8x0, but will that happen during its lifespan? At the moment, it doesn't appear so.
Given the DX11 marketshare and the number of Evergreens in developers hands....

And am I right in thinking all 5000 series cards have the same number of tessellation units ?
Redwood, Juniper and Cypress have the same per-clock geometry performance (setup and tessellation). Cedar is lower.

[Davros]

I have no real idea what redwood ect is stop using codenames ;)
what is cedar and why is it lower just clock speed ?

[3dcgi]

Cedar is the low end of the Evergreen chips. Radeon HD5400 I think. It's not just lower due to clock speed. Dave said "per-clock".