The NEXT LAST R600 Rumours & Speculation Thread

Well have a close look at the slides above and answer my question. I'm not debating the possible quality increase of a custom filter on top of Nx amount of MSAA samples one bit; in fact I'd say that we do need better filters (yet under presuppositions). I have a hard time digesting that the added filter samples get counted as AA sample densities.

Whether in the end the added filters deliver a tolerable grade of overfiltering for my taste or not, is something I have to decide when I see the method in real time.

My biggest objection with Quincunx and any related method is that the added filter isn't applied on just polygon edges/intersections, yet on the entire scene.

 

New 'advanced' custom AA resolve passes are nice and everything but it seems to me that nothing could stop competition to implement the same filters, in fact I'm willing to bet R6xx uses its shaders code to implement those filters.

 

Maybe their SFU is so close to being a full MAD that they just went the extra mile. It has 3 mini-MADs in it, actually:

Method and system for approximating sine and cosine functions

the biggest is 20-bit (just the mantissa). Dunno...

R5xx introduced a dedicated branch pipeline. The diagram appears to show a little blue runty thing, that's prolly it.

Jawed

 

They've been beefed up. We're just not sure how much yet. 4xAA per clock is looking very likely.

Jawed

 

Hmm, that way can we hope for Gaussian disperse kernel in future drivers?

Btw, what about that fully ring-bus topology -- how much it will differ from the current one in x1000 GPUs -- does that mean, that each ring-stop is now actually a 1/4 attached block of the whole design, a kind of on-chip distributed network?!

 

http://www.forum-3dcenter.org/vbulletin/showthread.php?p=5449558#post5449558

What's this?

R600:
64ALUs
16TMUs

RV630:
24ALUs
8TMUs

RV610:
8ALUs
4TMUs

 

Just about when I was happy that Quincunx was gone....heh

 

It's going to be much better than quincunx, I experimented a few months ago wiith custom AA resolve filters and even a simple gaussian filter on a 4X multisampled image (just a bit wider than a single pixel) was given extremely good results + a slight blur..

 

Blimey that AA is going to be fun to play with But only 2x, 4x and 8xMSAA? Hmm, I was hoping for 4x, 8x, 12x. I really want to see movies of these CFAA modes now...

I see I just about got there in time to post my speculation on the ALU operation

LOL, I moaned about the asymmetry of the ring-bus, now it's symmetric. I even convinced myself it was inevitably asymmetric, hahahaha.

I've asked before, I'm gonna ask again, what use are fp32-per-component bilinearly filtered texels?

Jawed

 

More ATI Slides, these showing performance comparison to Radeon 1950XTX, for Stencil Shadow, HDR, and Crossfire [ http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449740&postcount=414 ]

 

In theory, the R600 has slightly more math/shader power to throw at something like this, though.

Thanks, I missed that.

-Dave

 

Hasn't there been general talk for quite a while of moving away from hardcoded AA methods applied via control panel to the whole screen, to AA applied at the discretion of the developer directly where it is needed using shaders as the mechanism?

 

Can you share the results? (Pictures)

 

Well it's nice to see this thread got exciting again with at least rumors that R600 can beat GTX.

I still have a hard time believing it though, because why $400 if that's the case?

So glad they moved NDA up to May 2..

Someone back there mentioned supposedly Crysis running faster on R600..again that wouldn't surprise me if the TMU limits are true. Crysis is probably going to stress shaders like no game ever before by a factor of several.

 

That scale on the slides comparing it to the R580...
Shady, very shady.

 

That's true but you need also to sample your sub samples..

 

I noticed that, but it is somewhat mitigated by the R580 baseline being clearly marked, and the percentage increase clearly up the side, as well as just the frames on the bars.

 

And what would you have them compare it to? Or do you mean how they only show 1 block of "80 to 100%" and then leave 5 more blocks above that to go upto 200% ? Yeah, it's very marketting-like, but at least they provide you with raw numbers too.

 

I have no problem with that.
It's just that, this way, whoever sees those graphs and doesn't know any better will see a R600 with a bar that looks as if it was several times faster than a R580, when in fact it's closer to 0.8x times faster, at best.

If the graphs started at zero with 10% intervals, the impact would be much less pronounced.
I don't like it when there's this marketing effort trying to obscure clarity. Not everyone notices these subtleties.
And frankly, AMD is starting to "usurp" ATI's clear way of doing marketing (they took over this department directly, didn't they ?).

 

No objection there; one objection is that filter samples get counted as AA sample densities and a close second that if we'd be talking about 16x or even 24xCFAA how "slight" will the blur be there?