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

I don't imagine that AMD would want to maintain two distinct architectures for their GPUs, so yeah, VLIW4 sounds more reasonable.

 

Tahiti
New Zealand
Thames
Lombok

http://www.fudzilla.com/graphics/item/22503-amd-leaks-southern-islands-gpu-codenames

 

Tweaked VLIW4 would be my bet,for at least the $150+ cards. Maybe VLIW5 on Probably some attention to the dispatch, graphics set up stuffs, new generation tessellation unit, improved AF. More tweaking to memory controller for coalescing/simultaneous operations, as Cayman. POWER CONTAINMENT! Everything needs this. Everything.

 

Woot! I hope New Zealand is the dual chip architecture!

 

Thames? Never heard of that as an island

 

http://www.xbitlabs.com/news/other/..._Expects_Two_28nm_Tape_Outs_This_Quarter.html

Bolded are most likely the laptop editions

 

Ah that makes more sense.

Squilliam what if they do a Western Island instead & try to beat NV at their own game

 

Oh that would be nice!

 

Yeppers!

 

The end of the 64bit value parts (as far as the chips themselves are concerned, at least). Nvidia did it this gen, AMD next. Not surprising considering IGPs have 128bit at their disposal (albeit ddr3 only) of course. Still, with "only" 2x perf of Seymour Thames (5xVLIW4, 8 ROPs?) should be tiny (~70mm²) for a 128bit chip.

 

mczak, I think you missed the launch of GT 520/GF119....

 

Wow, I missed it too I think...baby Fermi.

 

Actually not, but I thought it's just the same chip as GF108/GF118 or whatever and just half of it disabled. That nvidia didn't publish any transistor count or die size numbers for it didn't help neither (and I haven't seen any measurements of die size anywhere). But maybe you've looked more closely at it and can tell me how many transistors it has .
So yeah if that's a different chip then nvidia hasn't abandoned separate chip for that segment yet - and worse it might be ddr3 only...

 

well gddr5 goes to waste on GF108 because of only four ROP (with 8 ROP as expected from a straight cut-down it would have helped with 4x and 16x AA)

the GF119 was said to be 96SP and 64bits, but geforce GT520 is 48SP only. so they can sell a full GF119 with 64bit gddr5 and get identical result to a GF108, but they have to sell those GF108 first.

 

Actually, techpowerup has measured die size - 79mm². So it is a different chip, and indeed probably only 48SP/64bit (it is only 60% the die size of GF108, after all).
You are probably right that GF108 wouldn't really have been much faster with gddr5 - nvidia chips also seem to be a bit less dependent on memory bandwidth than comparable AMD chips. Not sure though if GF119 supports gddr5 at all given the target market. The GT520 actually does quite well against a gddr5 HD6450, might be a draw (or it might even be faster) if both were equipped with ddr3. Of course, it has the usual 20% "die size penalty" against the amd chip, and raw performance probably isn't that relevant anyway.
It is strange though that nvidia completely skipped the low-end 64bit chip with the first wave of the GF10x chips, only to reintroduce it again with GF11x.

 

GT218 is a ~57mm2 DX10.1 chip - they probably felt that was sufficient for the ultra-low-end market in 2010, and it's still noticeably cheaper than what they can do with GF119. Interestingly, both GF118 and GF119 seem too small to be shrunk to 28nm - they'd very likely be pad limited with 64-bit and 128-bit memory interfaces respectively, unless you increased the SP count.

I wonder what die size AMD's Thames will have if it only targets twice the performance of Seymour. That looks like it'll be very close to the minimum die size for a 128-bit GPU as well then. I assume it's DDR3-only whereas Chelsea is 128-bit GDDR5, unless they are really the same chip. Also interesting that AMD is coming out with a native 192-bit GPU!

 

I thought features are an important selling point, even for low end. And GT218 is lacking there. Not to mention of course this one really is slow. And last I looked it didn't implement most of the quality-improving video features neither.
But yes, 79mm² vs. 57mm² means it's 40% larger. How much does that increase cost?
It also seems to draw (quite a bit) more power - of course it's also very significantly faster...
I'm just thinking it would have been a much more attractive option against the HD5450 - G210 had nothing going for it compared to HD5450 - less features, less performance,... A GF109 based card would have been quite a bit faster than HD5450. Don't forget nvidia released a half-disabled gf108 based gt420 (oem) - sure looks like there was some need for such a card.

Well for GF118 you could shrink it and move to 64bit gddr5 instead of 128bit ddr3 (though it means you can't get the oh-so-shiny 2GB low-end card - 4 2gbit gddr5 chips only give you 1GB and that already needs clamshell mode). I can't see a reason why you'd want to shrink GF119 since it would be slower than any new IGP, if you really need such a slow part just continue selling GF119 this time there's no new DX version it's lacking even.

Yes, I already wondered about Thames size too (a few posts above). I'd think though it will be gddr5 too, AMD had no problems supporting gddr5 this generation with the lowest end chip (of course that doesn't mean all implementations use gddr5) and I don't expect them to move back to ddr3. With twice the performance of Seymour you'd be very close to what Madison could do (which of course also supports ddr3 and gddr5). Seymour might use higher clocked gddr5 if that's a good idea in the mobile space.
The 192bit part is indeed interesting - not sure what to think of it yet...

 

thanks, I was starting to think the 96SP gf119 was wrong.
the power improvments from GF11x were much needed for the lowest end chip.

 

Outside GF100 perf/power increased maybe 10% so I can't see why they'd really depend on that for the low end chip. Idle power saw some more of an improvement (not sure how much of that really was due to chip changes) but at least for desktop would have been irrelevant for a low end chip anyway.

 

It's me or hardware news have been extremely underwhelming lately?

When can we expect a real new generation of cards? (either by Nvidia or AMD by the way).
Early 2012? It looks more and more like this year is set to be borrowing...