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

The pic's size of 600x535 pixels, if applied to a 365mm2 chip, gives roughly 880 pixels per mm2.
Each CU would then be between 6 and 7 mm2, maybe under 6.5.

In terms of CU area, almost half the die is in those arrays.
Almost half the die area in CUs and between 6 and 7mm^2 for each CU aren't both compatible with a die size of 365mm^2 :). CUs should be below 6mm^2. Estimate for CU size for Cape Verde was just 5mm^2 though that was using the blurry shot so possibly a bit off and Tahiti CUs are probably slightly larger (5% or so?).
 
I did some quick measuring and your hack is about 1.99x the size of Tahiti for a die area of ≈701 mm^2.

Haha! My pic is not meant to be taken to scale, since i stretched everything out to make it a square. Just double the yellow shaded portion (or increase it by 1.6x), and either double (or quadruple) the red shaded portion in fellix's diagram. Add that increased area to 252mm^2 and the result is probably around ~310mm^2, barring a nice rearrangement. These are just areas that I consider bottlenecks in AMD's design, although my opinion is in no way validated. Doubled TMUs is probably not needed, i just tried to throw that in there for good measure. I'm not an EE or programmer, merely a video game fan.:!:
 
Nice! Thank you, fellix. I hacked up your pic and doubled the ROPs and TMUs, quadrupled the setup engines and did 32-> 40 CUs : ^P Wonder how fast and big it would be? :LOL:
I don't think that makes a lot of sense. Quadrupling all the stuff in the middle seems entirely pointless, it already got 2 setup engines can't see why more than doubling would do anything at all for performance (plus that stuff in the middle probably has more stuff in there). I also think you doubled a fair bit more than just ROPs (e.g. MCs, though I can't figure out the ROP blocks individually). Even if not I don't think you'd gain much for more than 48 ROPs (unless you'd beef up the memory interface too).
 
Almost half the die area in CUs and between 6 and 7mm^2 for each CU aren't both compatible with a die size of 365mm^2 :). CUs should be below 6mm^2. Estimate for CU size for Cape Verde was just 5mm^2 though that was using the blurry shot so possibly a bit off and Tahiti CUs are probably slightly larger (5% or so?).

I did two separate calculations on the photo, and the overall error was smaller when calculating the total area of the array.
I was off by a few pixels in width when trying to measure the single CU, which has a bigger effect.
I tried zooming in a bit and trying to deliniate a CU, and it looks like the CU is smaller than what I measured the first time.
 
[FONT=verdana,geneva]Well, the R7870 used roughly a 100 Watts LESS then that GTX 580. When we reverse calculate and measure the power consumption the R7870 uses roughly 130 Watt where the GTX 580 hovers at a 235~240W, and that's measured in game while it's peaking and stressed. So for the R7850... well we measure roughly 106 Watt. Amazing stuff really.[/FONT]

Pretty impressive performance given the power envelope--I am curious how much memory power draw also plays into this. I do wish more people would offer some more settings in these reviews, e.g. BF3 with MSAA and w/o (ditto MSSAA+FXAA) would be nice. They mention 4xMSAA easily sucks down a ton of performance ("[FONT=verdana,geneva]4xA MSAA will cost you almost a third to half your framerate")[/FONT], so at 39fps for the 7870 at 1920x1200 does that work out to an average closer to 50fps or 60fps?
 
Ryan Shrout from PCPer
If there is a problem with today's review, it is that the cards aren't going to be in the market for two weeks - AMD expects market availability on March 19th. Why the review today then? Obviously AMD wants to be able to show off and discuss Pitcairn while the CeBit and GDC shows get under way this week and they realize they won't likely be able to contain leaks at both. In my view, AMD also wants to be sure that media had time to completely digest their 7000-series lineup before we head out to meet with NVIDIA to learn about Kepler - even though we would be under NDA it would surely affect our mindset knowing what is coming.
 
Amazing stuff you bring here @fellix and @jaredpace.. i think AMD may add 2 groups of CU quads(8xCUs) and it would look asymmetrical like GT200 and indeed mighty memory controller right there, no wonder ppl gets 8GHz memory speeds with heavy overclocks.. Pitcairn's and Cape Verde's controller looks just small..
 
Im a bit bored so im going to try and do a bit of analysis of Pitcairn and Cape Verde. Since the L2 cache of Pitcairn and CV are the same, the differences are mainly in the additional units (and of course 128 bit mem interface)

So compared to CV, Pitcairn adds 640 Shaders(10 CU's), 40 TMU's, 16 ROP's, a 128 bit memory interface and a geometry processing engine (did i miss anything?)

So all these things take up approx 2.8-1.5=1.3 billion transitors, and 212-123=89 mm2.

And given that the additional stuff is pretty much the configuration of CV, if we take CV's die, this means the L2 cache and the remaining fixed block(PCIE controller, Video decoder, VCE, Display outs, etc etc) take up 123-89=34 mm2. Now i dont know what to make of this number..and this is also a very simplistic analysis, the actual numbers could be a lot different..but like i said i was bored :smile:


So overall this kind of makes Tahiti seem a bit large in comparison to Pitcairn dosent it? Tahiti adds 768 Shaders(12 CU's), 48 TMU's, and an additional 128 bit memory interface (higher speed though), and takes 365-212=153 mm2 to do it all. Now i know the higher speed PHY's probably take up a lot more area, but it still seems like they werent as aggressive with transistor density on Tahiti as they were on Pitcairn. Pitcairn seems to have a bit over 10% greater transistor density than Tahiti.
 
Im a bit bored so im going to try and do a bit of analysis of Pitcairn and Cape Verde. Since the L2 cache of Pitcairn and CV are the same, the differences are mainly in the additional units (and of course 128 bit mem interface)

So compared to CV, Pitcairn adds 640 Shaders(10 CU's), 40 TMU's, 16 ROP's, a 128 bit memory interface and a geometry processing engine (did i miss anything?)

So all these things take up approx 2.8-1.5=1.3 billion transitors, and 212-123=89 mm2.

And given that the additional stuff is pretty much the configuration of CV, if we take CV's die, this means the L2 cache and the remaining fixed block(PCIE controller, Video decoder, VCE, Display outs, etc etc) take up 123-89=34 mm2. Now i dont know what to make of this number..and this is also a very simplistic analysis, the actual numbers could be a lot different..but like i said i was bored :smile:


So overall this kind of makes Tahiti seem a bit large in comparison to Pitcairn dosent it? Tahiti adds 768 Shaders(12 CU's), 48 TMU's, and an additional 128 bit memory interface (higher speed though), and takes 365-212=153 mm2 to do it all. Now i know the higher speed PHY's probably take up a lot more area, but it still seems like they werent as aggressive with transistor density on Tahiti as they were on Pitcairn. Pitcairn seems to have a bit over 10% greater transistor density than Tahiti.

I agree. So i think there will be a 7980 with transistor improvements ( and more shaders ) to fight GK110.
 
Tahiti implements 1/4th DP-rate, whereas both Cape Verde and Pitcairn go for the lowest possible 1/16th. Also, there probably is some fine grained redundancy in the larger chips in order to maximize yields which would make sense especially for the first parts on a new process.
 
According to chip marking Pitcairn is Made in China. Is this a new TSMC fab or were they forced to declare the actual production country?

edit:
There seems to be different Pitcairns around:
Taiwan Pitcairn @ Guru3D (bottom)
China Pitcairn @ Computerbase
Will it be the Fab in Shanghai? I know TSMC has a fab in Shanghai, probably established more than 5 years ago. But never had ita ppeared for mainstream chips...
 
Last edited by a moderator:
Tahiti implements 1/4th DP-rate, whereas both Cape Verde and Pitcairn go for the lowest possible 1/16th. Also, there probably is some fine grained redundancy in the larger chips in order to maximize yields which would make sense especially for the first parts on a new process.

Then, if they launch a 7980 with 1/16th dp-rate and with improved shaders (GCN 2.0 or whatever ) it should be not much larger than 7970 and have 6690 level performance. This would be a real R300 reborn !.

Because i have the feeling that GCN arquitecture as Rys said is very, very good, but has to be a little boosted to differenciate from the prior generation ( at least with the prices it has now ).

As now AMD has the architecture but no the "ATI 9700 pro" equivalent in consumer perception... although they are touching it with the finger tips...
 
Last edited by a moderator:
I also think that GCN is a major, major step for AMD. In every direction (well, maybe not backwards though) at once, it seems. But for an R300-level of epic the competition also has to fail with yet-to-be-presented-Kepler. :)
 
Back
Top