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

"In addition, the separate drive will also enable users to mix and match ATI graphics cards from different generations."

That pretty much means Crossfire between Trinity and GCN is possible, right? Provided the discrete GCN GPU isn't too fast relative to Trinity's, of course.

 

Mister Baumann, we've missed you: http://forum.beyond3d.com/showpost.php?p=1563991&postcount=23

But yeah, if we want to say that nothing has been officially said about the discrete segment, we can do that...for now

 

That's good news Dave. I think both AMD and Nvidia really need to step up on the dual gpu configs.

I did quite a bit of research over the past month because xfire/sli was something I was looking it since the price/performance can scale well.

However the following really turned me off to it:

- microstuttering
- scaling inconsistencies
- custom profiles
- general glitches and crashes


Ultimately, I decided to go with the fastest single GPU config because as real life takes priority, the time I do get for gaming, I don't want to spending on maintenance and optimization.

I'm glad to see AMD taking a large shot at it. I'll certainly re visit dual gpu solutions on my next upgrade (next fall) and hope it's very much "plug n forget."

 

The fact that no GPU can directly access the memory of the others in a multi-GPU system will always mean that there must be a software overhead that controls the duplicated frame-buffer data across the memory pools. No easy way out, really. I just hope that PCI-E 3.0 will ease to some point those issues.

 

the support of virtual memory in gnc and future nvidia will solve this problem?

 

Is this somehow different from memory virtualization? (which IIRC has been supported by AMD since HD2 series and was planned as DX10 feature, but canned (according to some rumors) from DX10 requirements in favor of nVidia since G80 didn't support it?)

 

Discrete GPU's will have this problem, even if virtual memory and coherency is extended over PCIe, as access to remote RAM over PCIe will be very slow.

If you duplicate the textures even in the presence of coherent virtual memory, you have reintroduced a sw layer in the middle.

 

7000 series delayed until early 2012:
http://www.nordichardware.com/news/71-graphics/44180-amd-hd-7000-delayed-to-2012.html

 

Megadrive1988: http://forum.beyond3d.com/showpost.php?p=1583244&postcount=760

 

Now I work in the business, I *FINALLY* get the "how can it be delayed?"

 

Interesting...
http://www.globalfoundries.com/eBooks/white papers/GF_Rambus_WhitePaper.aspx

So they are working on something together.

 

Ouch. Now Bergman is gone...

http://www.reuters.com/article/2011/09/22/amd-idUSL3E7KM40R20110922?feedType=RSS&feedName=companyNews&rpc=43

Wonder if this is just a shakeup or if things are really going that bad.

 

There's a gloom and doom thread in the 3D industry subforum for discussions on exec turnover.

Nothing so far seems to indicate that GCN is in serious trouble or at least in more trouble than anything else reliant on foundry success.
On the graphics side:
"Need input!"

 

http://semiaccurate.com/2011/09/22/amd-shows-off-28nm-gpus-and-trinity/

ahem....

 

I don't know how he's using the term MCM.

Is he saying the whole card is an MCM? That doesn't seem right, or at least it's pretty redundant to say a PCB with more than one component is an MCM.

Is he saying there is a package under that heatsink with two pieces of silicon? To that, I'd wonder if there would be a point.
There are MCMs where RAM gets put on the package, but it looks like there is RAM peeking out from under the cooler in that pick.

My brain just kind of puts out a NaN when given these inputs.

 

Did he mean MXM?

 

Very likely. It's a typo probably, C is just next to the X.

 

Usually when people say MCM they mean that multiple chips are directly bonded to a single substrate, not commonly the case for PCBs (since most ICs use interposers).

 

Anyone wanna guesstimate diesize of this thing? Looks to be at least "7800-series size", IMO

 

For me it's looking to be somewhat close to the Juniper size.