When is the next gen ATI/AMD Radeon due to hit? Northern Islands?

[Daedalus685]

Does anyone think ATI may try to release a RV870 derivative on 28nm first? (Just to test the waters)

Actually.. I think what we will get is the 6000 refresh at 28nm.
It is looking like no TSMC ability at 28nm until Q4.. so no tape out until the end of the year. At that schedule even if NI went straight to 28nm without a test we would be looking at Q2 or so before we saw the 6000 series (unless everything is perfect)... Given how much they loath the new arch with a new product thing that will be unlikely. The kind of time line that would allow 5000 series at 28nm first would put the 6000 series into late Q2 2011 at the earliest.... not likely.


I'd guess we will either see the rabbit out of the hat with ATI going with globalfo... or more likely the 6000 series at 40nm by the end of the year providing a relatively small jump in performance given the desire to remain at a small die. So we'd get a 6000 series that represents a small jump at first followed by a large refresh jump to 28nm.

 

[cbn]

Actually.. I think what we will get is the 6000 refresh at 28nm.
It is looking like no TSMC ability at 28nm until Q4.. so no tape out until the end of the year. At that schedule even if NI went straight to 28nm without a test we would be looking at Q2 or so before we saw the 6000 series (unless everything is perfect)... Given how much they loath the new arch with a new product thing that will be unlikely. The kind of time line that would allow 5000 series at 28nm first would put the 6000 series into late Q2 2011 at the earliest.... not likely.


I'd guess we will either see the rabbit out of the hat with ATI going with globalfo... or more likely the 6000 series at 40nm by the end of the year providing a relatively small jump in performance given the desire to remain at a small die. So we'd get a 6000 series that represents a small jump at first followed by a large refresh jump to 28nm.

If ATI goes HD6xxx on 40nm we are probably going to be talking large die right? Either that or mm2 for mm2 the new product is faster.

(Otherwise I can just hear all the folks saying they don't need the new architecture yet).

 

[Daedalus685]

If ATI goes HD6xxx on 40nm we are probably going to be talking large die right? Either that or mm2 for mm2 the new product is faster.

(Otherwise I can just hear all the folks saying they don't need the new architecture yet).

I doubt they will go large die honestly.. I think it is much more likely we see a generational gap that is not as large as we are used to.

It really depends what they add as far as whether it is worth it or not though. They could get 40% reasonably by removing some of the extra redundancy in the 5800 AT mentioned in the showing up for the fight write up coupled with the new architecture.

But who knows.. they will be silent on this for a while.. We may very well see Global Foundries making the thing.

Rumours over at S|A were stating that ATI had already taped out something (5890 if you will, maybe even the 6000 series) at 32nm before it was canned. With the APU being at 32nm at global that could provide some experience to make them risk the 6000 there.

 

[IntelUser2000]

The answer for what the node on next gen GPU seems to be quite simple. The first 40nm product from TSMC was in early 2009. Considering it takes average of 2 years to transition to the next full process node, it's unlikely they'll get 28nm this year. Plus the cancellation of their 32nm process doesn't seem to help either. We probably won't see 28nm until Q2 2011.

As for 28nm using GF process, that's also not likely, because designs for NI was probably already set few years ago, and you match your design to the process available. TSMC's process tech rules are probably much different from GlobalFoundries's 28nm. They are probably scrambling to get it ready on whatever is there, but it looks like 40nm now.

 

[GaiaHunter]

I doubt they will go large die honestly.. I think it is much more likely we see a generational gap that is not as large as we are used to.

It really depends what they add as far as whether it is worth it or not though. They could get 40% reasonably by removing some of the extra redundancy in the 5800 AT mentioned in the showing up for the fight write up coupled with the new architecture.

I agree with this.

ATI is supposed to redo their shaders (as most applications can't use all the 4 "less capable shaders" tacked with the fully programmable shader) and whatever else for next generation so it isn't going to be a "double all shaders RV780->RV870".

So there is the possibility for a 40nm 6000 series.

 

[Ares202]

So there is the possibility for a 40nm 6000 series.

I think the probability is high, what everyone should realise is that a new architecture could easily mean a smaller die size but much increased performance. I think they are going for a new architecture as their current design isnt scaling well at even 1600sp's (5850 clocked the same is only 2-3% slower)

I think HD 6000 will be released on Q4 2010/Q1 2011 and 28nm refresh in Q2/Q3 2011

Nvidia are in a very tricky position as their next move will have to be a slimmed down 32/28nm refresh of fermi with a few tweaks, which could be a full year away. Making the gpu any larger isnt really feasible on 40nm due to heat, power and cost constraints

 

[Mana]

Chances are we'll see a new generation of cards from both nVidia and ATI this fall. nVidia's only saving grace with the GTX 480 right now is that they have another team developing its successor with a likely launch period of fall 2010. Of course, ATI does too. Hopefully nVidia won't run into as many problems with the new card as they did the GTX 480.

 

[Rezist]

ATi made a larger die (mm wise) due to defects in the TSMC 40 nm node. nVidia was also affected by this and made a statement about it.

By the time the 6000 series comes out the defective 40 nm should be fixed and will give them some space for more XTOR's. As XTOR count has been increasing the performance isn't scaling as well, this is why you pretty much need a new node to get decent returns. Unless ATi engineers work some true magic it will be tough for the 6000 series to be much faster then the 5000 series at 40 nm.

This makes me think we'll just see a shrunk 5000 series once all the 40 nm issues are resolved and hold off on the 6000 series until 28 nm in Q2 2011.

I more or less expect the same from nVidia but they may wait longer due to a more complex chip for a refresh (getting the manufacturing perfected).

 

[Rezist]

Also I expect like the 4770 that as soon as TSMC can release 28 nm chips we'll see a 5790ish style chip at 28 nm But probably early Q1 2011.

 

[cbn]

A lot of you guys are saying ATI will not increase the die size on 40nm.

So who is going to buy the product then? Or will ATI use this architecture to perfect synchronized GPUs? (allowing dual GPU chips to feel like single GPU chips to the end user)

 

[Voo]

A lot of you guys are saying ATI will not increase the die size on 40nm.

So who is going to buy the product then?

Well I'd say that the 4890 was a rather nice card, so I don't see any reason why they couldn't do something similar again.

 

[GodisanAtheist]

A lot of you guys are saying ATI will not increase the die size on 40nm.

So who is going to buy the product then? Or will ATI use this architecture to perfect synchronized GPUs? (allowing dual GPU chips to feel like single GPU chips to the end user)

-Its just a question of efficiency. AMD's new arch could potentially do much more with much less, in which case sticking with 40nm could be feasible.

It could turn out like the R600 (or the fermi, for that matter), where the new arch is fundamentally quite sound it'd just need some tweaking to reach its full potential.

You do bring up a good point about the syncing GPUs. Seeing how ATI's whole strategy revolves around crossfire and scaling, we could see something completely different. Nvidia is pushing into the GPGPU space with a vengeance and their new arch demonstrates as much. ATI's long term goal is AMD's long term goal: Get fusion chips out there, and get people into a singularly AMD ecosystem (APUs on Fusion Core, getting discreet cards to crossfire with fusion increasing general system performance etc...).

 

[Nemesis 1]

Actually.. I think what we will get is the 6000 refresh at 28nm.
It is looking like no TSMC ability at 28nm until Q4.. so no tape out until the end of the year. At that schedule even if NI went straight to 28nm without a test we would be looking at Q2 or so before we saw the 6000 series (unless everything is perfect)... Given how much they loath the new arch with a new product thing that will be unlikely. The kind of time line that would allow 5000 series at 28nm first would put the 6000 series into late Q2 2011 at the earliest.... not likely.


I'd guess we will either see the rabbit out of the hat with ATI going with globalfo... or more likely the 6000 series at 40nm by the end of the year providing a relatively small jump in performance given the desire to remain at a small die. So we'd get a 6000 series that represents a small jump at first followed by a large refresh jump to 28nm.

.

Heres kinda how I see it. AMD buys ATI. AMD knows Fab processing + Those 45nm immerssion equipment came ALmost ready to go to the next shrink just minor expensive changes tho . GF has money. SO I am of the mind AMD was working on 32nm for long time than GF decided to stay with TSMC they wanted to go to 28 bulk also. I am thinking that Rather than scrape AMDs work at 32. They decided to Do it at 40nm. Give TSMC the contract at 40 . Than GF does the 28. and I am betting its out around jan1. But the 40nm what to do. When 28nm arrives Make the 40nm your mid range at less margin and ramp your high end at 28nm. Once thats ramped than Gf lets TSMC take the 28nm and produce it as AMD goes to the 22&20 nm . GF and AMD goal is to overtake Intel in processor Fab leadership. AMD / GF have to make the move befor 22. Or Its not going to happen . I think maybe . Its soooner than we think.

 

[Janooo]

It appears HD6000 won't be NI yet...

http://forum.beyond3d.com/showpost.php?p=1414576&postcount=414
http://forum.beyond3d.com/showpost.php?p=1414596&postcount=415
http://forum.beyond3d.com/showthread.php?t=55313&page=17

 

[Wreckage]

It will probably be out next year and compete well with NVIDIA's mid-range offerings.

 

[Nemesis 1]

Well Ya that makes since to me .

Copyed this.

Q3 listing HD6750, 40nm process, the chip area of more than Cypress, less than 400 mm2, still 1600SP/256bit bit wide, divided into two 800SP modules, each of which has an enhanced off Tessellation Unit and a Rasterizer, two modules parallel-oriented graphics. Triangles rate doubled, Tessellation 3-4 times performance improvement. L2 Cache redesigned to significantly improve the performance of GPGPU. Core frequency of 900Mhz-1GHz, TDP of 225 watts or so. Target performance is 10% -20%, GTX480, expectations and GF100 B1 version of its flagship chip, the performance was flat. North Island family, the first product in the maturity process of verifying the new structure.

Q4-Q1 next year, listing HD6670/6650, single 800SP modules / 128bit bit width, Water 28nm process.
Q1-Q2 next year, listing HD6870, 28nm process, four 800SP module, 512bit-bit wide return to core area of 400-450 mm between the target performance of dual-core card Fermi suppression. In fact, R600 is the ultimate form

 

[lifeblood]

Well Ya that makes since to me .

Copyed this.

Q3 listing HD6750, 40nm process, the chip area of more than Cypress, less than 400 mm2, still 1600SP/256bit bit wide, divided into two 800SP modules, each of which has an enhanced off Tessellation Unit and a Rasterizer, two modules parallel-oriented graphics. Triangles rate doubled, Tessellation 3-4 times performance improvement. L2 Cache redesigned to significantly improve the performance of GPGPU. Core frequency of 900Mhz-1GHz, TDP of 225 watts or so. Target performance is 10% -20%, GTX480, expectations and GF100 B1 version of its flagship chip, the performance was flat. North Island family, the first product in the maturity process of verifying the new structure.

Q4-Q1 next year, listing HD6670/6650, single 800SP modules / 128bit bit width, Water 28nm process.
Q1-Q2 next year, listing HD6870, 28nm process, four 800SP module, 512bit-bit wide return to core area of 400-450 mm between the target performance of dual-core card Fermi suppression. In fact, R600 is the ultimate form

Interesting, and makes sense. Bring out the upper mid-range card (6700) series on the old 40nm process to validate the architecture, bring out the lower mid-range card (6600) on the new process to validate the manufacturing, then bring out the high end card (6800) on the new process incorporating the lessons of both. It is different from the recent trend of bringing out the top cards first, but it makes sense given the schedule for 28nm.

 

[Lonyo]

Interesting, and makes sense. Bring out the upper mid-range card (6700) series on the old 40nm process to validate the architecture, bring out the lower mid-range card (6600) on the new process to validate the manufacturing, then bring out the high end card (6800) on the new process incorporating the lessons of both. It is different from the recent trend of bringing out the top cards first, but it makes sense given the schedule for 28nm.

The trend has never been for top cards on a full node I don't think. And typically when it gets attempted it goes horribly wrong. Half nodes are a bit different but still...
Ever since the days of the 9700/5800 it's been like that IIRC.
New high end card on new process = asking for trouble.

 

[Daedalus685]

The trend has never been for top cards on a full node I don't think. And typically when it gets attempted it goes horribly wrong. Half nodes are a bit different but still...
Ever since the days of the 9700/5800 it's been like that IIRC.
New high end card on new process = asking for trouble.

I think what people are saying is that what we will get would be akin to the 5770 being released first, then when 40nm came out the 5870 finding its way to market.

We would end up with a relatively small leap from 5870 to 67(70) but the 6000 refresh would take the form of the high end offering, or roughly double the performance.

We will have to wait and see what changed architecturally. But that makes the most sense... Instead of a 5870 refresh (we might still see a 5890) we will get a new architecture with a slight boost... then the new high end coming with a shrink.

The interesting thing is they seem to be looking at functional core modules for the next architecture... which could be interesting.

 

[GodisanAtheist]

Where the hell did nemesis come across that info?

It would be unfortunate to have AMD abandon the $200 "sweet spot" strategy and killer execution that has pretty much pulled its ass from the fire and start making all the same mistakes over again with big power hungry dies that inevitably balloon in cost and leave deadlines in the dust.

The 5 series already pushed it with the 360mm^2 dies, lets not head down the path of $400 MSRPs.

 

[Schmide]

Charlie on Southern Islands.

 

[Daedalus685]

Charlie on Southern Islands.

Most of that info is from some forum members at S|A. Though the southern island name is new.

I'm not sure I buy the "evergreen shaders" thing... NI is supposed to redo the 4+1 set up, and improve tessellation.. I'm not sure what they could to to tessellation if they use the same shaders... it would barely be a refresh.

I really think it is much more likely we see the True NI (or very close) in the fall, just the midrange part instead of the high end. When they can they will release at 28nm (who knows who's 28nm) with an entire product refresh and the true high end (now roughly the same die size as the mythical 6750).

 

[Schmide]

Did they say anything about side port?

 

[Daedalus685]

Did they say anything about side port?

Not directly.. though they were mentioning that the new architecture would be based off of 800 "shader" (the new ones most likely) modules. These would be the building blocks of the largest chip (which would use four of them)... That would have to sue something akin to sideport.

To say we know nothing yet is a vast understatement.. but I've figured for a while TSMC has a few more quarters of fucking our graphics cards in the ass... No shrink until 2011 = no great leaps in performance until then.

 

[Wreckage]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.