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

It will be interesting how pricing shakes out. Will NV peg back to AMD or will NV be aggressive and, if so, how will AMD ease in new pricing to compete? Unless there are major 28nm costs it seems AMD is in a position to reap the high prices now and move down in price while still having good margins later. I am sure that will be a delicate dance. Where and How they do so may have as much to say about the 7000-series in the end as the product itself (which is pretty amazing considering process node advancement).
 
I've just read Damien's review of Pitcairn. I'm impressed. I always expected it to be the killer chip in the GCN lineup, and I thought it would be faster than Cayman, but by 11%? That took me by surprise. Very nice chip, very well balanced, congratulations to AMD.
 
According to Anand, FP64 calculations are 1/16th rate as opposed to 1/4. What exactly have they turned off/cut out to achieve this?
 
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.
There's also ECC. Though if that's for L2 only it shouldn't make much difference but maybe it covers more.

I've just read Damien's review of Pitcairn. I'm impressed. I always expected it to be the killer chip in the GCN lineup, and I thought it would be faster than Cayman, but by 11%? That took me by surprise. Very nice chip, very well balanced, congratulations to AMD.
Yes, just like Barts was (imho) the killer chip. Surprising it's THAT small - certainly something below 250mm² was expected, but 212mm² is quite an achievement imho.
Maybe that was what nvidia was thinking when saying they expected gcn to be faster - if Tahiti had the same perf/area it would certainly be quite a bit faster. Based on raw stats you would expect the 7970 to be ~50% faster than HD7870, in reality it's more like 30% though there are indeed (mostly bandwidth-bound) games where it approaches 50%, but OTOH sometimes it's just ~10% faster. The overall balance of Pitcairn all things considered (die area, power consumption, memory bandwidth, CUs, ROPs, setup/rasterizer) seems to be quite a bit better (for gaming that is).

I'm relieved to see it does beat Cayman on average after all - guess those early print pcgh scores just were somewhat low due to having the 2 games (out of 7) where pitcairn does worst. All of the other reviews I've seen so far suggest the 7870 beats both GTX570 and HD6970 on average (though it's close sometimes).
 
Maybe that was what nvidia was thinking when saying they expected gcn to be faster - if Tahiti had the same perf/area it would certainly be quite a bit faster. Based on raw stats you would expect the 7970 to be ~50% faster than HD7870, in reality it's more like 30% though there are indeed (mostly bandwidth-bound) games where it approaches 50%, but OTOH sometimes it's just ~10% faster. The overall balance of Pitcairn all things considered (die area, power consumption, memory bandwidth, CUs, ROPs, setup/rasterizer) seems to be quite a bit better (for gaming that is).
I guess Tahiti is probably often limited by the scan-out throughput -- 2x16 fragments per clock may be coming a bit short for such a fat and hungry shader array. The peak INT8 texture rate also shows some deficiency (L2 interface?).
 
There's also ECC. Though if that's for L2 only it shouldn't make much difference but maybe it covers more.
ECC only makes sense, if it covers all R/W caches/memory and all read-only caches are at least parity protected. So ECC will most likely also cover the register files, LDS, and the vector L1 while the scalar L1 and instruction L1 caches probably use parity. I'm not sure about the ROP caches.
 
Ok, 4-3-3 times two for 7870, someone knows why the asymmetric design?

it should be more efficient than 4-4-2 -caches(K$,I$) wise..

gcn_cache_hierarchy7g1ii.png
 
Last edited by a moderator:
Isn't the HD7870 going to cannibalize the HD7950 sales, a lot?
How are they going to justify a $100 price premium for some 0-10% performance advantage?

BTW, this should be here:
ToWRT.jpg
 
Isn't the HD7870 going to cannibalize the HD7950 sales, a lot?
How are they going to justify a $100 price premium for some 0-10% performance advantage?
7950 GHz Editions? :LOL:

I wonder if in hopefully upcoming price wars a 256-Bit 2GB 5,5Gbps HD 7950 would be nice solution to reduce BOM, while reaching equal performance.
 
Should it? If it's not fake, it says GCN 2.0, so it shouldn't be Southern Islands but Sea Islands

No, Tenerife is one of the Canary Islands, which reside in the Atlantic Ocean.

For Sea Islands, you should be looking at islands that reside in seas, like Cyprus and Malta (Mediterranean Sea, just as examples).

So going by AMD's naming convention, this is in the right thread.
 
I have a distinct feeling that if my old 5750 doesn't off itself in the next month in my HTPC, it'll be declared dead regardless to justify a new 7850. ^^
 
Isn't the HD7870 going to cannibalize the HD7950 sales, a lot?
How are they going to justify a $100 price premium for some 0-10% performance advantage?
I think this is a very, very valid point. My overclocked 5850 is the very last "old" device in my new epic rig, and I've been strongly eyeballing the 7950 as the successor. After seeing the reviews of 7870, I don't know why I'd bother. A single 7870 would do everything I want and would cost less, and with such a tiny power draw, I could buy a second one later for even more punch.
 
No, Tenerife is one of the Canary Islands, which reside in the Atlantic Ocean.

For Sea Islands, you should be looking at islands that reside in seas, like Cyprus and Malta (Mediterranean Sea, just as examples).

So going by AMD's naming convention, this is in the right thread.

Ocean is just a big sea ;)

The naming conventions haven't been that "clear" before anyway, or how is "caribbean" considered "north(ern islands)"?
 
Back
Top