AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

[trinibwoy]

It makes no sense to complain that Tahiti should have more ROPs while it decisively beats a GTX580 in all fillrate benchmarks (and the GTX460 jaredpace mentioned is simply no contest).

Does it make sense to complain after comparing to Pitcairn?

Tahiti vs Pitcairn 1920x1080

Computerbase.de: 125%
Hardware.fr: 123%
Geometry: 93%
Pixel: 93%
Shader: 148%
Texture: 148%
Bandwidth: 171%

Even with an abundance of the big three (flops, texturing and bandwidth) Tahiti doesn't quite live up to expectations. So what's holding it back? ROPs, triangle setup, something completely different?

 

[3dilettante]

Tahiti was noted as having a higher performance hit when it comes to wide MRTs, which with certain settings leads to lower gains versus the GTX580 in BF3.
Adding an additional 50% in ROPs would appear to bump up the lower bound of the curve, such that the gains in BF3 and similar deferred engines would be higher in certain situations.

Potentially, other elements of the more strongly ordered graphics domain could also be bottlenecks, such as command processing, driver, GDS, and export bus.

 

[Gipsel]

Does it make sense to complain after comparing to Pitcairn?

If you put also Cypress and Cayman into the equation, it tells you that the ROPs are quite often not the predominant bottleneck.
The last sentence of of 3dilettante summarizes it well in my opinion, there is more than just setup rate, shader, ROPs, and memory bandwidth as potential bottlenecks.

 

[fellix]

Curious architecture comparison: Fragment rendering order - Fermi vs. Cayman vs. GCN

 

[Jawed]

That's so nice. I wonder if there's any meaning in the frame rate.

(Before I looked at the link I thought it was going to be the icare3d link that you posted in the comments.)

 

[fellix]

 

[Silent_Buddha]

Top to bottom GHZ lineup (for GCN). They were so close too.
I wonder if we will ever know why not?

I wouldn't be surprised if a Ghz baseline 7970 hit the same time as the GTX 680. Although I'm not sure it'd make much sense now that OC'd cards are so common.

Does it make sense to complain after comparing to Pitcairn?

Tahiti vs Pitcairn 1920x1080

Computerbase.de: 125%
Hardware.fr: 123%
Geometry: 93%
Pixel: 93%
Shader: 148%
Texture: 148%
Bandwidth: 171%

Even with an abundance of the big three (flops, texturing and bandwidth) Tahiti doesn't quite live up to expectations. So what's holding it back? ROPs, triangle setup, something completely different?

What's interesting about that is that Tahiti (comparing 7870 to 7970) also only has 92.5% of the clockspeed of Pitcairn. And if you follow normal rounding practices that means Tahiti has 93% of the clockspeed of Pitcairn.

That very closely matches the Geometry and Pixel numbers. But that makes absolute sense. Both Pitcarn and Tahiti share the same geometry and ROP front ends. Hence clockspeed will likely be the determining factor there when you aren't limited by other things (bandwidth, shaders, compute etc...).

So I suppose the only thing to complain about is that Pitcairn has a higher baseline frequency. Although I suppose if Tahiti had also launched at Ghz speed, that people would complain that in geometry and pixel that Pitcairn was the same speed as Tahiti. But at least in certain games, Tahiti would then have a larger advatage over Pitcairn.

Regards,
SB

 

[LordEC911]

Timing.

Although there was a relatively short time period between the releases of the chips, Verde and Pitcairn's bring-up, and to some extent qualification, have a reasonable level of leveraging going on so they are a little shortended in terms of initial engineering wafers back to product shipping. Actually setting the product "boundries" for Tahiti happened a while ago, on initial engineering material and few wafers out from the fab; Pitcairn and Verde on the other hand had their product boundries set when Tahiti production starts were already occuring and there is a very quick evolution in terms of understanding things with the new process / chips.

That makes sense. I find it so amazing what you guys are able to do with just a few production samples, in regards to adjusting future ASICs. Not to mention what you did with early samples of Tahiti to get it out ASAP and performing like it does.

I guess the question you want to ask is whether, now that we know things have evolved, are we going back to re-look at Tahiti.... ;-)

That was assumed but now that leaves more questions on your exact meaning... haha.

I wouldn't be surprised if a Ghz baseline 7970 hit the same time as the GTX 680. Although I'm not sure it'd make much sense now that OC'd cards are so common.

I had looked at that possibility as well but like you said, it makes no sense with the current AIB OC cards that are out there.
Though now we have Wavey's cryptic message about re-looking at Tahiti.

 

[Erinyes]

I guess the question you want to ask is whether, now that we know things have evolved, are we going back to re-look at Tahiti.... ;-)

Looks like you answered the question without anyone asking

The question is, are you looking only at Ghz or even higher?

 

[trinibwoy]

Potentially, other elements of the more strongly ordered graphics domain could also be bottlenecks, such as command processing, driver, GDS, and export bus.

Potentially, which means the current crop of theoretical benchmarks is quite unable to flesh out those nuances.

Whatever the cause, there's some interesting things happening. Looking at computerbase's numbers Tahiti gains no advantage over Pitcairn with 8xAA between 1680x1050 and 2560x1600. That's 2.3x the resolution so in theory it should pull away. Though with AA disabled you see more of a performance delta with increasing resolution.

 

[Rangers]

Side effect of 7870 being so dang fast and overclockable is, I doubt they'll have to drop price in response to gk104 in almost any case.

Bad side effect, i suppose.

 

[UniversalTruth]

Side effect of 7870 being so dang fast and overclockable is, I doubt they'll have to drop price in response to gk104 in almost any case.

Bad side effect, i suppose.

That's very strange. 7870 is not that fast, only 25% faster than GF114. Given the average performance improvements already seen, then we can (with 95% certainty) expect GK104 to beat Pitcairn handily, I mean to destroy it.
And if GK104 is priced at 350 or even 400$, who exactly will buy Pitcairn for 350$?

And then, I hope they don't think in the same way because that will leave many of us (you) still using (their) our good "old" cards.

 

[Rangers]

i'm talking about benchmarks like this

http://www.guru3d.com/imageview.php?image=37179

7870 oced beating nvidia 550 dollar flagship in almost every game, and pushing 7950/70.

if gk104 is 399, it might create a little pressure on 7870, any more, and none. if it's 499 for example, 7870 will probably feel no pressure.

Also, there is a question in my mind of gk104 performance, some seems to say usually faster than tahiti, some say trades blows, and some say a little slower than 7970. That's a pretty significant range.

In any case I am almost certain it will fall to 7970 at uber high res due to less bandwidth, even if normally the chip beats Tahiti, so there's also that.

 

[UniversalTruth]

I see your point.

Also, there is a question in my mind of gk104 performance, some seems to say usually faster than tahiti, some say trades blows, and some say a little slower than 7970. That's a pretty significant range.

The explanation I have seen is that those people are actually talking about different cards.

 

[Mianca]

In any case I am almost certain it will fall to 7970 at uber high res due to less bandwidth, even if normally the chip beats Tahiti.

Just look at trinibwoy's last post. Converting huge bandwidth advantages into hi-res dominance is basically one of HD 7970's most glaring problems. It boasts almost twice (!) the bandwidth of HD 7870 - yet it can't really pull away @ higher resolutions and quality settings.

I don't know what exactly is holding Tahiti back (I suspected ROPs until Gipsel made some interesting points in that respect) - but it doesn't exactly perform like it should compared to the way smaller, low-bandwidth Pitcairn.

For what it's worh, an aggressively clocked Pitcairn card paired with some fast memory could very well have launched in the same overall performance window as Tahiti - while still consuming less power

We know Tahiti is heavily underclocked, of course - but that still doesn't explain the conundrum around its poor bandwidth efficiency.

 

[caveman-jim]

I guess the question you want to ask is whether, now that we know things have evolved, are we going back to re-look at Tahiti.... ;-)

Why would you need to, sans competitive pressure?

Makes me think, AMD know how fast the GK104 could be and the re-work is already underway/finishing up (neglecting TSMC's 28nm production stoppage) for a response to GTX 680 launch; or, AMD don't and are waiting to see, and the GHz Edition 7970 won't be available anytime soon. If 7970 GHz Ed. is mid to late summer, early fall, will there be a 2012 8000 series? Will tweaked Tahiti be the 8000 series and Canary Islands pushed back until the next node? Can Dave really drink tea and eat digestive biscuits at the same time?

Inquiring minds want to know.

 

[3dilettante]

One thing they can review are the assumptions built into the DVFS calculations. If it turns out that the longer production run showed them to be too conservative, tweaking those can have affects on how high the design can be clocked and how aggressive powertune can be.

That might allow for a new edition of the card. The actual power control unit code might be a BIOS level update, which might allow for a reflash. If it were driver-updatable, it would be an unusually effective Catalyst upgrade, one superior to the "10% increase in benchmarked game X, everyone else kiss off" thing that happens now and again.

 

[Mianca]

Makes me think, AMD know how fast the GK104 could be and the re-work is already underway/finishing up (neglecting TSMC's 28nm production stoppage) for a response to GTX 680 launch; or, AMD don't and are waiting to see, and the GHz Edition 7970 won't be available anytime soon.

Given the alledged TMSC production stoppage and Tahiti's splendid overclocking headroom, I wouldn't be surprised if AMD just scrapped dual-GPU cards for the HD7*** series and gave us a "1.2 Ghz edition" (1Ghz won't cut it against already existing OC models) single-Tahiti-based HD7990.

Corresponding binning and testing has probably been under way for some time now.

 

[3dilettante]

AMD gives up ~90% scaling in a number of benchmarked games with crossfire on a stick for what will be at most 30% better performance?
This design would give up the bandwidth of a second bus and the additional memory, which is not wholly redundant.

 

[madyasiwi]

Makes me think, AMD know how fast the GK104 could be and the re-work is already underway/finishing up (neglecting TSMC's 28nm production stoppage) for a response to GTX 680 launch; or, AMD don't and are waiting to see...

Doesn't this image pretty much tells it all?