SemiAccurate: Southern Islands, Kepler, and Apple’s A6 process puzzle outed

[Jionix]

Nice little scoop by Charlie, if it proves to be true..

http://semiaccurate.com/2011/07/19/southern-islands-kepler-and-apples-a6-process-puzzle-outed/


This might be Cypress / Fermi all over again --- in terms of time to market.

 

[Crap Daddy]

And it will be the same regarding performance...

 

[Madcatatlas]

Well

1: if it arrives THAT long before Kepler, that they will also release a refresh at 28nm to compete against that, then it basicly wont have any competition.

2: the article is like Charlies normal spam, you cant really make an exact summary of what he is saying. Does he mean SI will come out on a low power capped process? in that case, wont the performance be lacking or not so good as a new process node entails?

I would if 2 is correct, assume 6870 or even 6950 like performance at half the power draw. Which in itself would be an insane design in terms of OEM and other market builds. Without competition, it would mean alot of sales for AMD.

 

[Jionix]

Probably. It depends on what chips AMD releases on this process. It's not meant for hungry, dense chips like Nvidia has been releasing. But AMD's light architecture seems able to be adapted into this process...

 

[Chinoman]

Yes, from what I'm reading I would expect small-die designs along the lines of R770 rather than the larger, almost Nvidia-esque Cypress.

 

[Jionix]

Well

1: if it arrives THAT long before Kepler, that they will also release a refresh at 28nm to compete against that, then it basicly wont have any competition.

2: the article is like Charlies normal spam, you cant really make an exact summary of what he is saying. Does he mean SI will come out on a low power capped process? in that case, wont the performance be lacking or not so good as a new process node entails?

I would if 2 is correct, assume 6870 or even 6950 like performance at half the power draw. Which in itself would be an insane design in terms of OEM and other market builds. Without competition, it would mean alot of sales for AMD.

This whole thing just smells of how AMD --ATI-- has been working the last couple years: Willingess to adapt. They have been very innovated in how they use their technology and how they pursue it, just like in the whole debacle of 40nm, and how ATI was able to find the secret sauce to get a good chip out of it is another example.

 

[notty22]

He laid out a bunch of scenario's there, typical. Even 40nm 7 series. He will refer back to this article and tell his minions how he was right. And that will be repeated and so on , etc ololo

 

[tijag]

And it will be the same regarding performance...

If this is true, its likely that AMD will also have a refresh of their product line launch on the 28nm HP process 6+ months later, close to when Kepler will launch.

This may be a situation where they can 'best of both worlds', with a launch now, and and an updated/improved board 6 months later.

IF that is what happens [big big IF], and IF AMD is launching a 79xx and a 78xx line, then their performance targets are probably lower than we had been guessing at.

7970 - GTX580 + 10%
7950 - GTX570 + 10%

7870 - equal to 6970
7850 - equal to 6950

7970 could cost less than Cayman to make, and they could charge enough to make a huge profit, all the while making Nvidia's pricing tumble down to remain competitive. Then in 6 months, when Kepler launches, they can go on the HP process as well, and do a refresh of the product line.

 

[Jionix]

He laid out a bunch of scenario's there, typical. Even 40nm 7 series. He will refer back to this article and tell his minions how he was right. And that will be repeated and so on , etc ololo

If you read the article thoroughly, he doesn't say that at all.

"The real question is what process they are going to make it on, the TSMC 40nm SiON or 28nm HKMG? 40nm would be big, hot, and limited, think volcanic island more than Southern, while the 28nm SHP HKMG process wasn’t supposed to be ready until late Q1, best case. The short story is that Southern Islands is very likely not on either one"

Obviously, he conditions it.. But, he specifically takes a stance on the topic.

 

[OCGuy]

I'm not even going to give him the page-hit....

 

[Jionix]

If this is true, its likely that AMD will also have a refresh of their product line launch on the 28nm HP process 6+ months later, close to when Kepler will launch.

This may be a situation where they can 'best of both worlds', with a launch now, and and an updated/improved board 6 months later.

IF that is what happens [big big IF], and IF AMD is launching a 79xx and a 78xx line, then their performance targets are probably lower than we had been guessing at.

7970 - GTX580 + 10%
7950 - GTX570 + 10%

7870 - equal to 6970
7850 - equal to 6950

7970 could cost less than Cayman to make, and they could charge enough to make a huge profit, all the while making Nvidia's pricing tumble down to remain competitive. Then in 6 months, when Kepler launches, they can go on the HP process as well, and do a refresh of the product line.

The big question I have is, has AMD been doing any test runs on this process? And if so, how long? Given how tight AMD is on leaks, this is the key and we might not know until they announce something

The ATI engineers figured out what was needed to make the early 40nm wafers work. If they have been doing test wafers here, I wouldn't put it past them to be able to figure out what it takes to get this node workable how they want it.

I think Charlie is using some some deduction here too. If anyone would attempt to use this as an ace in the hole, it's the graphics teams at AMD.

 

[darkswordsman17]

I'm not even going to give him the page-hit....

I don't know why anyone does.

 

[tviceman]

Probably. It depends on what chips AMD releases on this process. It's not meant for hungry, dense chips like Nvidia has been releasing. But AMD's light architecture seems able to be adapted into this process...

Nvidia has bigger chips, but AMD's chips are significantly more "dense" than Nvidia's are.... packing many more transistors per mm^2 than Nvidia.

 

[Keysplayr]

I quite literally have that domain blocked at my router. Could not see it if I tried. But I'm guessing that Charlie is playing the old "Lets see how much sh*t I can throw against the wall" game again. He can't go wrong!!!

 

[AtenRa]

I would love to hear what Idontcare have to say about the following photo,

GloFos 32nm xtor

What i really would like to know is the cons and negs for a Metal Gate shape like this. Is this a strain Gate or the shape has to do with the Gate-First in manufacturing ??

Intel 45nm NMOS (left)

High-K dielektrik in GloFos 32nm seams to be thicker

source :

 

[gorobei]

short version:
-tsmc 28nm hp hkmg is probably going to be late/buggy and cause delays in next gen.
-tsmc's 28nm lp (lower power transistors) are yielding just fine for xilinx and apple.
-amd's architecture isnt dependent on high frequency/high power, so they have the option to use 28 lp on southern islands now and switch to 28 hp when process matures.
-nvidia cant use 28 lp because they need higher clocks.

= amd can potentially deliver this year at 28 nm, while nv may be late or very late.

 

[badb0y]

short version:
-tsmc 28nm hp hkmg is probably going to be late/buggy and cause delays in next gen.
-tsmc's 28nm lp (lower power transistors) are yielding just fine for xilinx and apple.
-amd's architecture isnt dependent on high frequency/high power, so they have the option to use 28 lp on southern islands now and switch to 28 hp when process matures.
-nvidia cant use 28 lp because they need higher clocks.

= amd can potentially deliver this year at 28 nm, while nv may be late or very late.

Thanks for the summary, I don't want to dissect his article right now.

EDIT: So I took the time to read it since it wasn't that long and the nVidia bashing started on the last paragraph. He is saying Southern Islands is going to be a die shrunk 6000series particularly Cayman.

 

[ViRGE]

I quite literally have that domain blocked at my router. Could not see it if I tried. But I'm guessing that Charlie is playing the old "Lets see how much sh*t I can throw against the wall" game again. He can't go wrong!!!

I agree. You can't just port designs around like that, which makes the whole thing very far fetched. It's not impossible, but like every good crackpot scheme it has just enough truth to it that you can't immediately write it off.

AMD has never used a LP process for traditional GPUs. It's not just about clocks, it's about the size of the resulting chip and the inability to reuse functional units from other parts. To make this change would mean putting a lot of work into producing a one-off part that can't be much bigger than 100mm2 (which is small for a GPU), which would be in the neighborhood of Turks in size.

 

[gorobei]

there's no real mention of which architecture amd will use (gcn or vliw4), only that they have the option to deliver earlier than expected if they use 28nm lp rather than wait for 28 hp.

the chipworks blog link indicates that the 28 lp process requires some tweaking from the normal circuit structures and additional levels of redundancy(dummy traces).

 

[taltamir]

He laid out a bunch of scenario's there, typical. Even 40nm 7 series. He will refer back to this article and tell his minions how he was right. And that will be repeated and so on , etc ololo

ha,
"look, I made several vague conflicting predictions and one of them was right!"
charlie is such a tool.

 

[Lonyo]

Thanks for the summary, I don't want to dissect his article right now.

EDIT: So I took the time to read it since it wasn't that long and the nVidia bashing started on the last paragraph. He is saying Southern Islands is going to be a die shrunk 6000series particularly Cayman.

When I last saw a roadmap, the lower end was Cayman based VLIW4 and the higher end was the new architecture. It would seem not unreasonable that lower end VLIW4 based GPUs could be made on a low power oriented process which isn't ideal for large/high density chips, assuming AMDs (current) architecture will play nice.
Also makes sense to put the low end on the process which makes them as small as possible to improve margins, probably.
Wouldn't be the first time they have a non-top end chip on a new process either (HD4770), and NV did it with 40nm IIRC (low end 200 series were their first 40nm chips).
So it's not unreasonable to believe that some of SI might be on a different process and launch earlier than other chips at 28nm, but who knows (apart from AMD and TSMC).

 

[Borealis7]

These pictures look like my dental work...

in any case, i don't see how this is good news for us Gamers, since now we can't really expect that the 7000 series will bring performance increases if its released early.

 

[3DVagabond]

These pictures look like my dental work...

in any case, i don't see how this is good news for us Gamers, since now we can't really expect that the 7000 series will bring performance increases if its released early.

It should bring at least price drops, even if performance isn't substantially better (Which nobody has shown that it won't be substantially better.). That's alone is not a bad thing for gamers. Only for nVidia if they have to compete with 40nm parts.

 

[Silverforce11]

Normal LP chips including dual cores powering smart phones typically run @~1ghz, usually not much higher. Thats exactly matching the speed AMD's GPU run at. While NV does indeed go >1.6ghz on parts of their designs. Charlie is saying AMD is using a modified 28nm LP, not the regular type. Who knows what will happen. It sounds very plausible given AMD's willingness to try/pioneer new technology and also fit the other rumors concerning delays in HKMG until 2012.

SA is really the only source of info since he obviously has contacts in TSMC.
Everyone else has to wait until AIBs get their hands on samples for leaks.

 

[Jionix]

Nvidia has bigger chips, but AMD's chips are significantly more "dense" than Nvidia's are.... packing many more transistors per mm^2 than Nvidia.

You're right, I worded it wrong, as I meant bigger, but neither dense nor bigger sounds like thety would be suited for this process..

AMD could repeat last years release strategy --- mid-range first again. Whether they focus on power and heat, or go for as much performance as the process will allow..