AMD: R7xx Speculation

[Jawed]

Jawed: your shader to resolve a 4X render target can probably be much simpler, assuming bilinear filtering is enabled it can be carried on in one clock cycle (and just one tex2D instruction)

The other side of the coin: can the RBEs provide more than 16 pixels per second of samples? Additionally, the colour fillrate is 16 pixels per second, so an MSAA resolve shader that runs in only one cycle doesn't actually run any faster than the shader I've described.

RV635 has one quad RBE (4 pixels per clock). A TU-based MSAA resolve would run at 8 pixels (32 samples) per clock. My MSAA resolve shader would run at 6 pixels per clock. Both techniques are bottlenecked by the single quad-RBE.

RV620 can perform 4 pixels (16 samples) per clock through the TUs, matching the RBEs. The MSAA resolve shader I wrote would run at 1 pixel every 2 clocks. So in this case TU-bilinear averaging would be fast enough while my shader would run at half-rate.

So, I wonder if there are any benchtests of RV620 (RV610) 4xMSAA. The behaviour of this in comparison with higher spec GPUs might provide some clues as to the basic form of MSAA resolve. If bandwidth doesn't get it first...

Jawed

 

[Kaotik]

AFAIK default quality is the same as HQ (default is the same as HQ on R5xx anyway). The only thing that might cause shimmering AFAICS is having CatAI set to high.

What I'm wondering is how did ComputerBase get "AF" numbers on Radeons, since the CCC doesn't allow to choose between HQ (default) and "normal"

 

[Lukfi]

I think ATI Tray Tools can do that, perhaps they used it? Btw, ninelven, where did you learn that Catalyst AI on High turns on those "performance over quality" optimizations?

 

[Jawed]

AFAIK default quality is the same as HQ (default is the same as HQ on R5xx anyway). The only thing that might cause shimmering AFAICS is having CatAI set to high.

Hmm, I was under the impression that R5xx IQ genuinely benefits from HQ AF.

http://www.techreport.com/articles.x/12458/5

Ignoring the D3D AF tester pix, it's quite clear that R5xx at default AF quality is inferior.

Since that article was published the AF options for R6xx appear to have changed from "you have no choice" to "default or HQ". I don't know if, after this change, default equals the IQ of the older driver.

Jawed

 

[ninelven]

Hmm, I was under the impression that R5xx IQ genuinely benefits from HQ AF

It does; I was saying the default R6xx quality is the same as the R5xx HQ.

I'm not sure why a HQ option exists for R6xx products... perhaps it forces HQ AF even when CatAI is set to high?

Btw, ninelven, where did you learn that Catalyst AI on High turns on those "performance over quality" optimizations?

My own personal experience. But a site did a nice article on it a while back... can't remember which one though. I'll see if I can dig it up.

 

[no-X]

What I'm wondering is how did ComputerBase get "AF" numbers on Radeons, since the CCC doesn't allow to choose between HQ (default) and "normal"

Older drivers allowed to choose it.

ninelven: Advanced AI performs some kind of frequency analysis of textures (on CPU) and on the basis of the result suitable filtering is applied. It should result in better looking textures at the expense of higher CPU load.

 

[ninelven]

It should result in better looking textures at the expense of higher CPU load.

Funny, all I've ever recieved is nasty shimmering.

 

[CarstenS]

What I'm wondering is how did ComputerBase get "AF" numbers on Radeons, since the CCC doesn't allow to choose between HQ (default) and "normal"

As no-X mentioned older drivers - in fact i only remember launch drivers - could have allowed that.

But one could read into german text underneath that with "HQ" they actually mean AI switched off (and that'd explain much of the sudden drop in performance even with only 2x AF)

My dear colleague did also some IQ-Tests on the Radeon with newer drivers and found, that turning AI off does not improve IQ as much anymore as in the time of the R5xx... not even for those with newer cats, as a matter of fact.
http://www.pcgameshardware.de/aid,603118/News/Qualitaet_der_anisotropen_Filterung_R600_gegen_R5x0/
(I am sure you all can put this into your favourite translator - google, for example.

 

[trinibwoy]

http://www.computerbase.de/artikel/...d_3870_rv670/5/#abschnitt_aa_und_afskalierung

Jawed

Thanks. Much prefer B3D's presentation though. English is just a bonus

Though these graphs show that any reviewer forcing HQ on AMD's parts isn't doing them any favors.

 

[Jawed]

As no-X mentioned older drivers - in fact i only remember launch drivers - could have allowed that.

But one could read into german text underneath that with "HQ" they actually mean AI switched off (and that'd explain much of the sudden drop in performance even with only 2x AF)

My dear colleague did also some IQ-Tests on the Radeon with newer drivers and found, that turning AI off does not improve IQ as much anymore as in the time of the R5xx... not even for those with newer cats, as a matter of fact.
http://www.pcgameshardware.de/aid,603118/News/Qualitaet_der_anisotropen_Filterung_R600_gegen_R5x0/
(I am sure you all can put this into your favourite translator - google, for example.

Thanks.

With various AI settings R6xx IQ varies less than R5xx, but in terms of shimmer R600 is never as good as R5xx at its best?

AI only affects sharpness/shimmer (and performance). Angle-dependency is not affected by AI, is it?

Jawed

 

[A.L.M.]

Hi everyone,
look at what I found just doing some link guessing on the PowerColor site?

 

[LordEC911]

Heard that often already, but many things never made it down the road

I personally still think it was defective, whatever they say. Noone in their sane mind would have wanted a slower AA implementation in a new chip.

Slower than what? It was faster than R580...

Its required for DX10.1, no?
I wonder if NV choosed to stay 10.0 partially because of this - ie they keep it simpler for them and higher speed, lack 10.1, but as they dominate the market they can easily make it look like another ps1.4 vs ps1.3 till DX11

Was it not suppose to be in DX10 spec until MS changed it for Nvidia?

 

[Karoshi]

GIGABYTE HD3950 512MB / 256 bit

http://www.hexus.net/content/item.php?item=13600

Cant remember any references to a 3950.
Googling brings up lots of typos. But dont think its a typo, both model number GV-R395-512H-B and marketing sticker have the 3950 reference.
What is it?
Higher clocked (respin?) 3870?
Recycled SP-impaired rv7x0?

 

[XMAN26]

Slower than what? It was faster than R580...

Was it not suppose to be in DX10 spec until MS changed it for Nvidia?

I believe he is referring the the 2900XT being slower than the 1950XTX in certain cases for quite some time after launch.

 

[ZerazaX]

GIGABYTE HD3950 512MB / 256 bit

http://www.hexus.net/content/item.php?item=13600

Cant remember any references to a 3950.
Googling brings up lots of typos. But dont think its a typo, both model number GV-R395-512H-B and marketing sticker have the 3950 reference.
What is it?
Higher clocked (respin?) 3870?
Recycled SP-impaired rv7x0?

Good question - I always wondered if those Rev A12 RV670s were ever supposed to be released, but who knows if they're still going with it or if they're just cut-down RV770's

Should be interesting - looks like its supposed to be slotted between the 4850 and 3870

 

[Kaotik]

By any chance RV770 with 320 SPs but other enhancements included (Excluding perhaps 32 TMUs)?

 

[ZerazaX]

BTW looking at that link, I see how Gigabyte has the Sp count for the GT200's but not the ATI cards...

Weird to think about that, so either ATI is hiding something still or they have something they don't want to let the competition know

 

[CJ]

Hmm. Interesting. I've been leaking some NV stuff lately... I think it's time to leak some AMD stuff too... how about a Vantage X-number of R700?

 

[digitalwanderer]

You wouldn't dare!

 

[trinibwoy]

Hmm. Interesting. I've been leaking some NV stuff lately... I think it's time to leak some AMD stuff too... how about a Vantage X-number of R700?

Heh, sure why not. The internet erupting in flames is always a good show