AMD: R9xx Speculation

[GZ007]

I don't see how a dedicated HW would be justified for such a "loose" AA method, that is no more different than the tons of other post-process image effects. The edge-detect CFAA doesn't have it, why should MLAA?

Maybe because its several times faster than MSAA which couldnt be said about the custom AA modes.

http://www.iryokufx.com/mlaa/

 

[mczak]

Maybe because its several times faster than MSAA which couldnt be said about the custom AA modes.

http://www.iryokufx.com/mlaa/

I have some issues with this. How can it be several times faster than MSAA? I've never seen MSAA taking a more than 10 times hit (!) as they are claiming (and if MSAA only has like a 50% hit it's hard to be "several times faster"...). Are they comparing this specifically in situations where the MSAA solution runs out of memory?

 

[Gipsel]

I have some issues with this. How can it be several times faster than MSAA? I've never seen MSAA taking a more than 10 times hit (!) as they are claiming (and if MSAA only has like a 50% hit it's hard to be "several times faster"...).

They don't compare the absolute timings per frame. Instead they use the additional time needed for the AA (frame time with AA - frame time without it).

 

[jimbo75]

So how does that work then? Take a game like Crysis Warhead which has an 8xMSAA penalty of around 30% fps on a 5870 (compared to no AA), does that mean this hardware morphological AA would only have a penalty of 3-4% fps?

 

[ShaidarHaran]

Because you can't just "remove" the T unit without somehow compensating for its specific functionality?

True, but there are several ways to do that, one of which would be to add functionality to two of the remaining units and allow them to work simultaneously on a problem that would've previously been delegated to the fat unit. Keeping two narrow units and two slightly wider units should still result in a smaller ALU than the previous organization, IMHO.

(c) how many AMD engineers sold their souls to the perf/mm² devil to actually achieve (at least) HD 5850 performance @ just 60% of Cypress' die size?

The largest graphics bottlenecks in today's games are bandwidth and math rate, two things which Barts would match or surpass Cypress so why shouldn't it be ~ Cypress performance despite being smaller due to the new ALU layout? I don't know that Barts will be 60% of Cypress though, that does seem low. I could see Barts at 70-80% of Cypress' die size achieving 90-95% of the performance of Cypress.

 

[Kaotik]

Isn't the current EdgeDetect already a form of MLAA?

 

[fellix]

Not really, because MLAA is a pure image post-process method and it can't recover sub-pixel information -- one of its biggest drawbacks.

 

[Mintmaster]

512-bit sounds way overkill for Cayman, doesn't it? I mean, Cypress was fine with its 256-bit bus and 4.8GT/s memory. Cayman can most likely rely on 6~7GT/s, so… maybe 384-bit, but 512?

Cypress wasn't really "fine". The 4890 was 15-50% faster than the 5770, despite having the same clock speed, ROPs, and SPs.

"Smaller than Cypress" and "better perf/mm²" is to be expected - I just don't believe in Charlie's <200mm² number.

I wouldn't expect that. If you adjust for process node, NVidia has been getting worse per mm2 since G92. Well, maybe GF104 was a step up from GT200, but they're pretty close. IIRC, ATI took a step back with Evergreen (again, adjusting for process node).

Then again, there's no major features being added (except maybe multiple tris per clock - please!), so I suppose it's not unreasonable to expect improvement.

 

[Gipsel]

Fudo claims that Barts is a minor update and shaders will be similar to the Evergreen series in the comments of his last bit about the HD6800 launch.

It is [a] minor [refresh], shaders are similar, just a few more of them.

At least it would end the confusion about the organization of the SIMDs with seemingly 80 SPs each (32 would be an alternative ).

 

[ZerazaX]

Fudo claims that Barts is a minor update and shaders will be similar to the Evergreen series in the comments of his last bit about the HD6800 launch.

At least it would end the confusion about the organization of the SIMDs with seemingly 80 SPs each (32 would be an alternative ).

Fudo's been pretty off on the ATI front the last few years I believe

 

[Bouncing Zabaglione Bros.]

Fudo claims that Barts is a minor update and shaders will be similar to the Evergreen series in the comments of his last bit about the HD6800 launch.

At least it would end the confusion about the organization of the SIMDs with seemingly 80 SPs each (32 would be an alternative ).

Could be disinformation - Fudo does tend to get fed by Nvidia's PR teat.

Hard to believe the reveal is now only a day away. AMD certainly knows how to keep a secret nowadays.

 

[Rebel44]

Hard to believe the reveal is now only a day away. AMD certainly knows how to keep a secret nowadays.

Yeah, I remember days when we would know complete specs of new HW month before release.

 

[Alexko]

Fudo's been pretty off on the ATI front the last few years I believe

Fudo's been pretty off on everything for a while. He does tend to get near-term future AMD CPUs right, though, for some reason.

 

[Sontin]

Fudo's been pretty off on the ATI front the last few years I believe

He was very accourate with the performance increase from cypress over rv770. More than other guys...

 

[rpg.314]

Fudo's been pretty off on everything for a while. He does tend to get near-term future AMD CPUs right, though, for some reason.

Like that's hard. Just say, modest improvement, but still slower than competition.

 

[Alexko]

Like that's hard. Just say, modest improvement, but still slower than competition.

No I mean he got precise SKU specs and launch dates. But yeah, it was never really anything unexpected.

 

[Broken Hope]

The 6800 series drivers have surfaced.

http://www.station-drivers.com/tele...yst_8.782-vista7(www.station-drivers.com).exe

[ATI.Mfg.NTamd64.6.1]
"Radeon HD 6800 Series" = ati2mtag_NI, PCI\VEN_1002&DEV_6738
"Radeon HD 6800 Series " = ati2mtag_NI, PCI\VEN_1002&DEV_6739

 

[Unknown Soldier]

Fudu says that the 6870 is officially 900Mhz core and 1050MHz GDDR5 memory.

http://www.fudzilla.com/graphics/item/20486-radeon-hd-6870-officially-at-900mhz

He also says the official launch date is 22nd Oct.

http://www.fudzilla.com/graphics/item/20485-radeon-hd-6870-and-6850-launch-oct-22nd

 

[TKK]

The 6800 series drivers have surfaced.

http://www.station-drivers.com/telechargement/ati/catalyst/ati_catalyst_8.782-vista7(www.station-drivers.com).exe

[ATI.Mfg.NTamd64.6.1]
"Radeon HD 6800 Series" = ati2mtag_NI, PCI\VEN_1002&DEV_6738
"Radeon HD 6800 Series " = ati2mtag_NI, PCI\VEN_1002&DEV_6739

And for anyone who still doubts that 6800=Barts, rememver these entries can be found in the drivers (forgot which file) as well:

243,BARTS XT (6738),NI BARTS
244,BARTS PRO (6739),NI BARTS

 

[Kaldskryke]

Not sure this is worth posting, but pcper has written a short speculation article here. His take on the VLIW-4 arrangement seems particularly strange to me... isn't it the t-unit that is supposedly underutilized? Why go to 3+1?