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

Yeah it'll be my choice of GPU once they come out for the Sager/Clevos

 

Here I thought you were asking an 'easy' question. Back when DH posted the exact same roadmap (without a pretty slide) in December, they claimed Wimbledon was planned be 25% faster than Blackcomb. That is approx. the exact same difference as 6870->6970 (desktop). It's also about half what Mobility 6970 is over 5870. A 256-bit part will retain 32 ROPs, and I doubt the memory speed will scale much higher (at most up to 5Gbps from 3.6Gbps, probably ~4gbps), so one could kind of extrapolate the similarities in performance from the desktop for mobility, slightly adjusted for clockspeed differences. Hence, Wimbledon will probably be a cut-down shrink of something similar to Cayman on 28nm; somewhere around 1280sp. I would not be surprised if Wimbledon is essentially a '5830-like' part relegated to the mobile space.

You're not asking that question though, you're asking if the next enthusiast chip will be at least 14% faster than 6970 (which is 14% faster than 5870 on avg). I'd say the answer is a definitive yes; because it has to be a better alternative to GTX580 to make sense. I'd say the baseline necessity is at least ~25% overall.

I'm of the opinion AMD will shoot for ~4TF with greater than 6Gbps GDDR5; enough to put them ahead of GTX580 in both raw performance and bandwidth. It may not end up being their largest chip on 28nm, but I do believe it is where they will begin. Who knows if that will be called 7970, 7870, or something else. Also, who knows what combination of shaders/clockspeed it will take to get there. Feasibly, I could see anything from 1792-2048 making sense depending upon if they continue to go the high-voltage route or rather opt for more logic at a power-friendly clock speed. I haven't the slightest. There has been rumblings 28nm has shown to be more about power savings than clock improvements, which would suggest more logic, but that could all be bullshit.

 

Yeah, you're right i guess. I was asking the question because it's quite hard to envision (for me at least) how exactly AMD is going to achieve this on 40nm (without increasing die size, but i would almost bet against this). Especially since they had allegedly already did an "optimization round" which resulted in the 6970.

But maybe they'll strike when we least expect it; just like 3870->4870.

 

It seems those couple of month old fake specs have surfaced again.

Those are trivially fake: shader and TMU counts don't match.

 

I wonder when fakers will bother to learn such basics…

 

From the R9xx thread about Southern Islands

 

Is it taped or tapped out? I always got confused over this serious issue...

 

So he basically says, the design is so far ahead of the process technology that it is hard to predict the retail product availability. It can be anytime between August 2011 and August 2012.

 

Or maybe on the flip side the process technology isn't as far behind the design as previously thought? Does this make sense? Maybe they are indeed going to use GF 28nm fabrication on this and the typical rumour mill hasn't adjusted to this change in supplier model for GPU production and therefore the supporting process rumour hasn't come along to point to an eventual production date even though it isn't too far into the future?

Im just speculating here, feel free to shoot down my reasoning but im just a little cautious with GPU rumours given the possible entrance of another player in this market.

 

I guess for those who are somewhat following the manufacturing side of things it's pretty clear that a retail GPU on 28nm this year is close to impossible. My feeling is even Charley agrees with it throughout his piece.

 

AMD to launch three new Radeon HD 6000 series cards on April 19

Report: Radeon HD 7000 series about to hit mass production
28 nm? Don't think so.

 

A 28nm pipe cleaner might come first. It could be a low volume part like 4770.

 

Could be on Global foundries 32nm process.

My thought on specs...
32 SIMDs(2048 VLIW4 units) with either 2 or 4 set up engines, 128 TUs, 32 ROPs. (SI pro being - 28 SIMDs(1792 VLIW4 units) divided by 2 or 4 set up engines, 112 TUs, 32 ROPs).

 

As far as I know, GF doesn't have 32nm bulk process

 

But AMD is already making Evergreens (with some K10 cores attached ) on 32nm SOI.

 

GPUs with CPU-Socket compatibility - finally!

 

Really?

 

A 28nm pipe cleaner = a 28nm Barts.
The question is v5 or v4?

 

Depends on whether the product is one that has been on the roadmap for quite some time or if it is a reaction to the lateness of TSMC's 28nm process. If the former, I would expect it to be a VLIW-4 architecture, the latter of course being a VLIW-5 architecture.

We know now that Cayman was a 40nm "backport" of a 32nm design which belonged to the original Northern Islands family of GPUs, but was of course delayed due to the cancellation of TSMC's 32nm process and the subsequent delay of the 28nm process.

 

a vliw4 part makes more sense to me.