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

Their problems in pro market aren't related to mem bw and would not be fixed in any way whatsoever by throwing wider bus at Tahiti.

 

While it's possible that it might have escaped notice, I don't remember you arguing for it before.

 

Kepler is a full node ahead of GF100, so you can expect a FLOPS increase close to 2×. The FLOPS/bandwidth ratio has been continually increasing over the last generations, so I wouldn't expect the same increase in bandwidth, but it should still go up by quite a bit to avoid severe bottlenecks.

Plus, running GDDR rated for 6GHz at this very speed is not easy, it requires a lot of effort on the memory controller and PHY—an area where NVIDIA doesn't particularly shine—and can be relatively inefficient, power-wise. So a 512-bit bus doesn't sound unrealistic at all.

Not that I have any inside information or personal experience of this, just my 2 cents.

 

R600 graphics core was quite a bit smaller than the die size alone suggested. The memory interface padding at the perimeter was mostly stacked and took a generous portion of the IC area, which is the reason for the "miraculous" fitting of 512-bit interface in just 420 sq.mm at 80nm process. A very unbalanced design, later beaten by RV670.

 

Start here:

http://forum.beyond3d.com/showthread.php?p=1569124#post1569124

I'm sure there are other times too, but I got bored rummaging.

This all comes back to this:

http://forum.beyond3d.com/showpost.php?p=1485272&postcount=3870

Plus we also know that a primary goal of GCN is to solve the scaling gotchas.

 

Well if Cypress' MC is almost twice the size of Redwood's just to move from 4.2Ghz to 4.8Ghz and we have no idea how much larger Cayman's MC is to hit 5.5ghz maybe they thought it would be much easier to up the bus size and use a smaller/simpler, though slower, MC to get the bandwidth increase they wanted.

 

I worried about the memory limit as well. 7xxx series should required more bandwidth with a new design of SM.
a 384bit MC must be helpful.

 

Rambus - AMD's new secret weapon against Nvidia!


http://translate.google.com/transla...170-rambus-amds-neue-geheimwaffe-gegen-nvidia

 

When I google for XDR2, the only real, non-white paper products I find are Michelin tires.

I hope it stays that way.

 

I' am actually looking forward to see XDR2 in action compare to GDDR5 on Radeon GPU.

 

Pulling random numbers out of thin air, it would appear. Let's stop doing that. Another example is Anand's claims about the MCs in Redwood and respectively Cypress, which are bogus (only some bits of the MC actually abide that relationship, not it as a whole) - yet the Internetz is running with it with great speed, it would appear.

 

ah you mean this (bottom of the page)

 

Actually XDR2 couldn't be used in Low-End cos the price.
And different MC may cause more money in redesign compared to a flexible desgin

 

Maybe AMD is going with XDR2 because they had to design a XDR2 memory controller for PS4 GPU...

 

i'm not for the xdr theory, but highend and lowend gpu haven't different mc yet?

 

If the 7xxx series is going to use xdr2, where are they going to get the ram from? I mean who is going to produce it? as apparently no one has even produced a sample of xdr2 yet. Or has 7xxx been moved back to late in 2012?

 

In my opinion, this whole "XDR2 for HD 7000 high-end" is BS.
It probably started as an enthusiast's speculation/wishful thinking/wet dream, was picked up by a site that of course didn't name any source, then other sites picked up that story for teh clickz, and now many of the enthusiasts among their readers apply the it's all over the net and I want it to be true, so it must be true™-filter and defend it vigorously.

My personal I learned the hard way that such rumors usually turn out to be a pile of goo™-filter tells me we'll probably see at most a 384-bit controller with mildly increased GDDR5 clocks on Tahiti. Best case.

But we won't see anything fancy until one of the following generations introduces an interposer + stacked memory.

 

It's like a desgin with modularization now.

Exactly