AMD: R8xx Speculation

[nAo]

It's a little more work to really calculate perspective correct differentials for texture coordinates per pixel instead of the difference between pixels, but it's by no means impossible.

Exactly, and this is what properly implemented deferred shading engines do. On the other hand you end up paying the cost of explicit derivatives (i.e.samplegradient()), which usually run at 1/4 throughput (afaik).

 

[trinibwoy]

Express what you really want to say about that

Can't really say much until you know what that test is actually doing. In general Futuremark has been very vague about the synthetic tests in Vantage, rendering them even more useless.

 

[Andrew Lauritzen]

Exactly, and this is what properly implemented deferred shading engines do. On the other hand you end up paying the cost of explicit derivatives (i.e.samplegradient()), which usually run at 1/4 throughput (afaik).

I'm not sure that in the general case of computing derivatives for arbitrarily parameterized/computed values it turns out to be any cheaper than just running the shader (at least to that point) on at least two adjacent pixels and computing the differences.

In any case I doubt the 2x2 thing will change any time soon in any of the current graphics APIs.

 

[Ailuros]

Perhaps it's not worth to have dedicated unit, when it's desired usage leads to inevitable inefficiencies in other areas of hardware...

Tesselation == checkbox feature, to satisfy DX certificate?

I can't know what they've done exactly but I can't imagine a ff tesselation unit to capture that much die area either.

 

[Shtal]

Based from this chart, it looks to me GTX285 is inferior tech. compare to HD5870.

EDIT: ATI nailed Nvidia with RV870.... Perfect example of what ATI did with ATI Radeon 9700Pro vs. GF4Ti 4600

 

[Voxilla]

Based from this chart, it looks to me GTX285 is inferior tech. compare to HD5870.
EDIT: ATI nailed Nvidia with RV870.... Perfect example of what ATI did with ATI Radeon 9700Pro vs. GF4Ti 4600

Makes, me laugh. The R600 was also very inferior to the G80.
Comparing previous tech to new tech isn't very convincing either.

 

[psolord]

@shtal

Isn't the 5000 series supposed to go against GT300 series and its derivatives?

================================

Some independent benchmarks from one lucky XS member:

first verification pic
http://filesmelt.com/downloader/DSC00205.JPG

now some scores,

HD5870 stock - i7 965 stock (3.2GHz)

3DMark06 - 16xAF forced in CCC
22383 3DMarks
SM 2.0 Score 8704
SM 3.0 Score 10655
CPU Score 6282

No AF forced in CCC
22549 3DMarks

DMC 4 1920x1080 8xAA 16xAF
scene 1 162.78
scene 2 123.86
scene 3 221.67
scene 4 118.59

resident evil 5 1920x1080 max details 8xAA
97fps

Thanks for bringing this to our attention!

My lowly GTX 260 can already do RE5 1920X1080+4XAA at 94.8fps! Ok 4XAA is different from 8XAA, but so is 5870 and GTX 260!

Granted, in DMC4, both my 4870X2 or my GTX260, were not nearly as impressive. I really hope we will see more DMC4 scenarios than RE5 ones! :S

 

[Shtal]

Makes, me laugh. The R600 was also very inferior to the G80.
Comparing previous tech to new tech isn't very convincing either.

The reason I said that, is because Nvidia seemed aggressive and always ahead in the game and what now worries Nvidia is that they cannot do instant reply to RV870 as they always did after the days of ATI R300.

EDIT: And this is just approx where Nvidia is with GT300 : http://www.semiaccurate.com/2009/09/15/nvidia-gt300-yeilds-under-2/

 

[Voxilla]

The reason I said that, is because Nvidia seemed aggressive and always ahead in the game and what now worries Nvidia is that they cannot do instant reply to RV870 as they always did after the days of ATI R300.

EDIT: And this is just approx where Nvidia is with GT300 : http://www.semiaccurate.com/2009/09/15/nvidia-gt300-yeilds-under-2/

Lately ATI was always in advance with finer process technology.
If say R870 is 2 billion transistors and GT300 is 3 billion on the same 40 nm process,
I don't see a reason why yield would be dramatically worse as it is proportional to die size.

Architecturally GT300 will probably leap into new territories where R870 should be similar to R770.
Also a 512 bit bus is quite a bit more challenging than a 256 bit one.

 

[no-X]

list of DX11 games...

compilation:

• Alien vs. Predator: DX11, tesselator, Q1/10
• Battlefield: Bad Company 2: DX11, Q1/10
• BattleForge: DX11, direct compute, via patch, 09/09 (thx2 Demirug)
• Crysis 2: DX10.1, DX11
• Dirt 2: DX11, tesselator, 12/09
• Dungeons and Dragons Online: Eberron Unlimited: DX11, TBA
• F1 2010: DX11, 2010
• Genghis Khan: DX11, TBA
• Lord of ther Rings Online: DX11, Q1/10
• Race Driver: Grid 2: DX11
• Stalker: Call of Pripyat: DX11, 11/09

• Frostbite 2 Engine: DX11, 2010
• Vision Engine: DX11, TBA

 

[neliz]

no-x BC2 will use Frostbite 1, not FB2. Can't assume DX11 compatibility

 

[Demirug]

Should be:

BattleForge: DX11, Direct Compute, via patch, 09/09

 

[hoom]

I'm still wondering if they did anything new to boost the maximum triangle rate of small triangles.

2* rasteriser isn't enough for you?

Real geometry via displacement maps, even if a big performance hit, looks really nice compared to POM or normals, and if you are running at nearly 100fps anyhow why not take the hit for some real geometry?

Hey, who needs raytracing & software rendering (see Epic guy rants/Larrabee) when (mostly?) FF tessellation can give you sub-pixel triangles where it counts already?

I dare you to compare those two pictures and say "Enh, POM looks as good", because it simply doesn't.

I'm not arguing with that since I see tessellation as the biggest leap forward since HDR (or even SM2.0?!) but there is definitely some tessellation artifacts there.
Especially up on the roof ridge-line you can see where triangles dip down visibly.
Its interesting that there was no Tessellated Wireframe pic of that angle (maybe the OP forgot to post that pic?)

Giving bricks volume is fine but they already did that in Toyshop with POM.

The edge bricks were geometry (see the wireframe mode), POM only did the flat surface & that falls over at high AoA & edges/silhouettes while Tessellation doesn't because its real geometry.

 

[no-X]

neliz: Sources:

http://www.pcgameshardware.com/aid,...efields-Frostbite-Engine-already-ported/News/
http://en.wikipedia.org/wiki/List_of_games_with_DirectX_11_support

Demirug: Thanks.

 

[hoom]

Yes tesselation can be scaled depending on the framerate. It is parameter under complete control of shader programmer.

So a new benchmarking paradigm (was going to use that yesterday but lost the word half-way through posting ) it is then!

Benchmark tests will have to be per AA/AF level still but with the tessellation level graphed over time rather than FPS since FPS should remain at close enough to constant 60FPS.

Demirug do you have comment on that point, being an actual dev working on a DX11 title?

 

[2senile]

I don't see a reason why yield would be dramatically worse as it is proportional to die size.

.

This was touched on lightly in the GT300 thread. Yield could be dramatically different due to design & other factors. In fact, yields for two parts, same area & transistor count, could differ widely.

I'll even hazard to guess that the current success of GT21x does not bode well.

I only have a faint idea how this stuff works so hopefully we'll have some more knowledgeable posters filling in the details.

 

[Demirug]

Demirug do you have comment on that point, being an actual dev working on a DX11 title?

I am the lead engine engineer for BattleForge.

 

[Voxilla]

2* rasteriser isn't enough for you?

Cool, that is a good start !

I'm not arguing with that since I see tessellation as the biggest leap forward since HDR (or even SM2.0?!)

It is also long overdue as it originally already was announced for DX9.

 

[neliz]

Should be:

BattleForge: DX11, Direct Compute, via patch, 09/09

Nice hint there

 

[flynn]

Lately ATI was always in advance with finer process technology.
If say R870 is 2 billion transistors and GT300 is 3 billion on the same 40 nm process,
I don't see a reason why yield would be dramatically worse as it is proportional to die size.

I was wondering about that, unless there's something I'm missing I see no reason for NV's yields to be so ridiculously low vs ATI's.

However, if they need 50% more transistor budget and a bus twice as wide to complete with ATI's design they're doing something wrong.