fermi has dedicated Tesselation hardware

[josiahsuarez]

according to Chris Ray

http://www.rage3d.com/board/showpost.php?p=1336134574&postcount=1573
So.., what happens in vegas stays in vegas... that being said. The NDA isn;t that far off from coming off for another technical layout of "Graphics" Fermi. So I won't be able to talk to freely about it until then.

However, I can say this at this point as I expect alot of leaks in the nearby future. Fermi dramatically improves on of icecold's biggest complaints about Gt200 and really all cards are having this problem. I expect he'll be happy to see it addressed GF100's tessellation engine is impressive. Its certainly not software and there certainly is dedicated tessellation hardware beyond just the HULL/Shader routines. Its been an incredibly long couple months of me being quiet. But thats at an end really shortly.

the problem he's talking about is GT200 setup engine capable of only doing one triangle per clock

[Madpistol]

If that is true, then that would explain why GF100 clobbers the 5870 on the Unigen Heaven Benchmark.

That being said, that means that the GF100 probably isn't much more powerful than the 5870, except for the implementation of dedicated tesselation hardware. GF100 just got really interesting.

[Johnny C]

If that is true, then that would explain why GF100 clobbers the 5870 on the Unigen Heaven Benchmark.

That being said, that means that the GF100 probably isn't much more powerful than the 5870, except for the implementation of dedicated tesselation hardware. GF100 just got really interesting.

Does it clobber the 5870? Links? (lee)

[FlakMagnet]

@Madpistol : Are you saying that the ATi cards perform tessellation in software and that's why Fermi's hardware implementation is better?

Would be really surprised if the ATi cards didn't do tessellation in hardware. But saying that, I am surprised Fermi doesn't do it using it's 512 processing cores. If it did, it would allow nVidia to tweak the routines to make it better/faster by a simple driver update.

P.S. I thought the fact that Fermi has dedicated tessellation hardware was a known fact for some time now...?

[newparad1gm]

I'm not too familiar with this, but does this mean that the 5870 does its tessellation as one of the operations that it processes through its stream processors, alongside vertex and pixel shading, while the GF100 has dedicated hardware specifically for tessellation, and is unused when tessellation does not occur? Sort of like how the XBOX 360 GPU has a daughter die with 10MB eDRAM specifically for z-buffer operations and anti-aliasing.

[DansFace]

the 5870 is a direct x 11 chip... doesn't it also do Tesselation in hardware?

[DansFace]

http://www.guru3d.com/article/radeon-hd-5870-review-test/7

"ATI already had a hardware tessellation unit in their GPUs for years. But the older units could not be addressed whatsoever in DirectX. The tessellation units featured in the ATI Radeon HD 2000, HD 3000 and HD 4000 series are all very much based on the same functionality found in the XBOX 360 'Xenos' graphics chip."

[josiahsuarez]

yes the 5870 et al have dedicated tessellation hardware also. but there has been some confusion about whether GF100 would do it in software, at least to me it appeared there was some confusion.

[DansFace]

lol...
are we going to see something like the FX5800 and how well it handled directX 9 shaders :D

[Maverick123w]

lol...
are we going to see something like the FX5800 and how well it handled directX 9 shaders :D

The FX5800 was awesome in Doom 3 but I think that was about it. It was god awful in dx9 for sure.

[Ninja Prime]

according to Chris Ray

http://www.rage3d.com/board/showpost.php?p=1336134574&postcount=1573


the problem he's talking about is GT200 setup engine capable of only doing one triangle per clock

Who cares? The reason no one has raised the triangle rate is there is no need to. No game comes even remotely close to maxing the triangle rate of any card since like Geforce 4. I think even Crysis maxed out peaks at like 2 million polys a frame, at 60 fps (which you can't run maxed Crysis at anyway) even clocked at only 600 mhz you have 10 million polys a frame. 5 times as much as needed, for the most demanding game out there.

[jlippo]

Who cares? The reason no one has raised the triangle rate is there is no need to. No game comes even remotely close to maxing the triangle rate of any card since like Geforce 4. I think even Crysis maxed out peaks at like 2 million polys a frame, at 60 fps (which you can't run maxed Crysis at anyway) even clocked at only 600 mhz you have 10 million polys a frame. 5 times as much as needed, for the most demanding game out there.
yes, but with something like tesselation triangle counts go into whole different level of sillyness.

The biggest problem with the bottleneck is that the rest of the GPU just waits for the results doing nothing, not very good use of the ALUs.
If the GPU can skip invisible or setup polygons faster than one polygon per clock, it will help in cases like rendering shadow buffer of a highly tesselated scene.
Also if it is combined with ability to render more than one polygon at the time, we might see some interesting numbers. ;)

[Redeemed]

I'm not holding my breath for Fermi.

However, I really do want it to be awesome. Right now I'm expecting to pick up at least one of the HD5xxx series refresh cards. However, if Fermi delivers without breaking the bank I will consider one of them as well.

My dual 8800GTXs suffice, for now, but I'm planning on purchasing a larger monitor meaning I'd no longer be gaming at 1680x1050 but 1920x1200. My dual 8800GTXs are just barely able to push this res with most newer titles, I know they'd struggle at 1920x1200 big time. :(

[nekrosoft13]

the 5870 is a direct x 11 chip... doesn't it also do Tesselation in hardware?

i think the point is that

5870 does it in the GPU or VPU as ATI likes ot call it.

and Fermi will have extra dedicated chip for tesselation, leaving the GPU to do other functions.

[Madpistol]

@Madpistol : Are you saying that the ATi cards perform tessellation in software and that's why Fermi's hardware implementation is better?

Would be really surprised if the ATi cards didn't do tessellation in hardware. But saying that, I am surprised Fermi doesn't do it using it's 512 processing cores. If it did, it would allow nVidia to tweak the routines to make it better/faster by a simple driver update.

P.S. I thought the fact that Fermi has dedicated tessellation hardware was a known fact for some time now...?

No. ATI's Tesselation is definitely hardware accelerated. Otherwise, any DX10 card could do it, and we all know that the only cards on the market currently capable of tesselation is AMD's HD 5xxx line, which are inherently DX11.



This video is a demonstration of Fermi @ CES 2010. I'm not sure if this is the Dual chip or single chip solution, but you can make that decision for yourself.

gkI-ThRTrPY

If you watch, it is incredibly smooth, both with Tesselation on and off. The resolution is either 1680x1050 or 1920x1200 (based on measurements of the options in the top left of the screen, I think it's 1680x1050) and if this video was recorded @ 30FPS, then we know that the demo is running at at least 30FPS. However, based on what this guy is saying, it might even be 60FPS.

What I do know is that @ 1680x1050 on my HD 5870, this is the results I get from benchmarking w/ Tesselation on:

http://www.nvnews.net/vbulletin/attachment.php?attachmentid=39144&stc=1&d=1263420535

The settings are default; DX11, tesselation enabled, anistropy 4x, all high settings. The only difference, is that the resolution is 1680x1050.

Now, here's some of the low FPS points that I observed throughout the demo:

Dragon spin: 19 FPS minimum
Full roof: 22 FPS minimum
street walk: 22 FPS minimum
Wall and Stairs: 25 FPS minimum.


The only section that I didn't see in the video above was the Dragon and the Full roof, which ironically are the most geometry intensive parts of the demo. From what I did see, though, Fermi is running the entire demo ridiculously well. If Fermi doesn't have a dedicated "Tesselation Shader," then nvidia optimized the demo for CES, or Fermi is 2-3x as powerful as an HD 5870.

Take it as you will, but that's crazy. :headexplode:

Bear in mind that when he activated the wireframe mode, the framerate slipped to about 30-45 FPS. The HD 5870 slips to about 10-20 FPS in heavily tesselated areas.

Last thing. The tesselated parts of this demo are at or above CGI level graphics for triangle count. There's probably close to 10 million triangles on the Dragon alone.

[nekrosoft13]

what case did he say? maingear?

[Madpistol]

what case did he say? maingear?

Yep. It's called the Maingear Shift.

http://computershopper.com/desktops/reviews/maingear-shift

[Johnny C]

I hate to say it, especially since my 965BE is on it's way, but I think that your Phenom II is holding back your results.

I've seen results with an average 50+fps @ 1920X1080 same settings you with a 5850....That 5870 should be rocking Heaven better than that.

[Sazar]

Yep. It's called the Maingear Shift.

http://computershopper.com/desktops/reviews/maingear-shift

The screen looks to be the same as what I have, which is a Dell 24" ultrasharp, so it is likely running at 1920x1200.

By the way, where are you seeing the frame-rate? I looked and looked and couldn't find a shot where it was legible.

[Redeemed]

Okay.

Awesome.

Fermi is 2 to 3 times faster than a single HD5870.

In one benchmark application.

How about some real in-game numbers? Crysis? Crysis WH? FC2? Or any other more recent titles?

Until these are available, I'll reserve judgement.

[Slytat]

Okay.

Awesome.

Fermi is 2 to 3 times faster than a single HD5870.

In one benchmark application.

How about some real in-game numbers? Crysis? Crysis WH? FC2? Or any other more recent titles?

Until these are available, I'll reserve judgement.

Ditto.

We need some RW numbers from a reliable source (AKA Not Fudzilla).

[Madpistol]

I hate to say it, especially since my 965BE is on it's way, but I think that your Phenom II is holding back your results.

I've seen results with an average 50+fps @ 1920X1080 same settings you with a 5850....That 5870 should be rocking Heaven better than that.

Link?

I'd believe you, but my CPU usage maxes at 40% on this demo. On the really high triangle points, it's only a hair above idle.

Either that, or Phenom II's just suck at this demo. It wouldn't surprise me either.

[Madpistol]

The screen looks to be the same as what I have, which is a Dell 24" ultrasharp, so it is likely running at 1920x1200.

By the way, where are you seeing the frame-rate? I looked and looked and couldn't find a shot where it was legible.

I'm not seeing the framerate. Just watch the video. The demo is running incredibly smooth. Smooth = 50-60FPS, BUT since the video is being shot @ 30FPS, it could be lower.

Basically, I used common sense.

[Redeemed]

I've long since gave up on synthetic benches. Give me in-game numbers, then I'll care. ;)

[Viral]

I've long since gave up on synthetic benches. Give me in-game numbers, then I'll care. ;)

+1