ATI's next gen process details [S|A]

[StrangerGuy]

Same reason why there are only 480 for the release version of the GTX480.




Can we all agree that TSMC is blowing it for us?

Having 3 billion transistors on a immature process of a second rate silicon manufacturer is like purposely asking for disaster.

 

[VirtualLarry]

What about the alternative possibility? THat the silicon wafer picture thought to be GF's 28nm bulk process, is really a wafer of ATI GPUs, and that they are already qualifying GF's 28nm process?

 

[jvroig]

Having 3 billion transistors on a immature process of a second rate silicon manufacturer is like purposely asking for disaster.

You consider TSMC a second rate silicon manufacturer? The TSMC who owns 50% of the foundry market, 80% of the profits, and considered as a first source of semiconductor manufacture?

I know people are pissed at how 40nm turned out at TSMC, but calling TSMC second rate is going a little too far.

 

[StrangerGuy]

You consider TSMC a second rate silicon manufacturer? The TSMC who owns 50% of the foundry market, 80% of the profits, and considered as a first source of semiconductor manufacture?

I know people are pissed at how 40nm turned out at TSMC, but calling TSMC second rate is going a little too far.

Never heard of Intel?

 

[NoQuarter]

Honestly this falls more in line with Intel's much better tick - tock strategy. New architecture on same die size, then die shrink, then new architecture on same die, then die shrink. That way you take less risks.

 

[cbn]

Honestly this falls more in line with Intel's much better tick - tock strategy. New architecture on same die size, then die shrink, then new architecture on same die, then die shrink. That way you take less risks.

I think I would rather see midrange N. Islands on 40nm with sideport than this S. Islands scheme (half new design) that is being talked about.

 

[NoQuarter]

I think I would rather see midrange N. Islands on 40nm with sideport than this S. Islands scheme (half new design) that is being talked about.

Oh, I didn't read all the details. Yea I don't know about doing a hybrid. I'd rather see a midrange N. Islands on 40nm too.

 

[HurleyBird]

If this is true, then ATI's engineers are being very, very aggresive. Lets see: They started RV870 as a more complex design, and redid it halfway through due to TSMC's screw ups. Despite this they get their entire next gen line up out before Nvidia ships their first DX11 card, and Nvidia's ASIC performs only about the same. Then TSMC screws up again so they decide to make an entire new hybrid family of cards on 40nm in addition to NI. At the same time they are working on Llano 32nm SOI, the graphics IP for snapdragon, probably new graphics cards/IP for Nintendo and MS next gen console, and the graphics for the Windows 7 phone based off the Xbox360 GPU. Wow, Let all that sink in for awhile. Assuming that this is what AMD is doing, and assuming that they actually pull all of this off, then this will probably be the biggest engineering feat/streak that the industry has ever known. Has ATI bitten off more than they can chew? And if not, how the heck is this the same company that put out R600 only a few years ago?

Also, if the ATI team can pull all this off, I think they're wasting a lot of energy. There is no real need to update Evergreen so quickly -- the family is more than competitive. They'd probably be better off splitting into two families/teams, one that only aims at gaming performance, and the other that only aims at stream performance. Fermi is a hybrid that fails at game performance (as an architecture in terms of die size and power consumption), but is almost definitely a better stream processor than the competition (at least in more complex scenarios where it's efficiency can beat out evergreens monster theoretical FLOPS). A family of processors aimed at only attaining high stream performance could do the same thing to Fermi in GPGPU that Evergreen is doing to Fermi in graphics. If you have the engineering resources for so many projects, you should try and elimate some of the overlap.

 

[Dribble]

Be very surprised if Ati have an *intermediate* architecture. Why fix what isn't broken - the 58xx series is working fine. They don't have a problem in the gaming market, they have a problem in the gpu compute market and an intermediate architecture isn't going to fix that.

 

[at80eighty]

It sounds like Charlie is trying to sugar coat that there is something wrong with Northern Islands.

there there, Scott

 

[Will Robinson]

If the 28nm mode is looking a little shaky the smart play is to stay with 40nm and see how some of these new generation parts perform while using an architecture that is well understood as the framework.
Currently configured, the 5850 has a fair bit of headroom for higher clocks.
With faster Vram and NI performance improvements it could make for a decent Cypress refresh without risking the whole enchilada on the new node.

 

[Madcatatlas]

I do think ATI needs to rethink its next gen card. As to what actually IS their next gen card? thats still in the air i think.

It will float down in the coming months im guessing

 

[Lonyo]

I do think ATI needs to rethink its next gen card. As to what actually IS their next gen card? thats still in the air i think.

It will float down in the coming months im guessing

So you think they need to rethink their next gen card, even though you don't know anything about their next gen card?

 

[Madcatatlas]

your so smart i was just reading my own post and thinking, do these words look right..hmm!

I ment: they need to think about going away from the previous 4 years of gpu design. Even more so than Fermi was ment to

 

[Daedalus685]

I really find it hard to imagine they will have a real 'intermediate' design...

I find it far more likely that we either get refresh that has better tessellation (sort of a refresh prime) or they actually introduce NI at 40nm, but call it SI because it is at 40nm..

One thing I am almost certain of at this point... NI is 28nm, we will not see its true form until 2011, and ATI will not be releasing a fermi level GPU area (a la 2900) just to get it out on a bigger node. Which is good.... I'm still sceptical about all this though, looks like a boring year is coming

 

[yh125d]

Same reason why there are only 480 for the release version of the GTX480.




Can we all agree that TSMC is blowing it for us?

Absolutely



The rumors for SI/NI sound kinda promising, I'm digging the modular design bit (kind of more akin to CPU manufacturing, 2-6 big core modules linked by shared caches/uncores/stuff like that) but who knows if the 2x800 part rumored will be enough to combat a fermi refresh. I doubt it will do better than match the current fermi, and NV I suspect is working on a refresh... a 725mHx 512sp B1 Fermi would be a wonder to behold, and could possibly keep thermals manageable and not cost too much either. ATI'd have a tough battle fighting that one


Any rumors around regarding NVs plans? (fermi refresh, next gen, 28nm)

 

[Shilohen]

If this is true, then ATI's engineers are being very, very aggresive. Lets see: They started RV870 as a more complex design, and redid it halfway through due to TSMC's screw ups. Despite this they get their entire next gen line up out before Nvidia ships their first DX11 card, and Nvidia's ASIC performs only about the same. Then TSMC screws up again so they decide to make an entire new hybrid family of cards on 40nm in addition to NI. At the same time they are working on Llano 32nm SOI, the graphics IP for snapdragon, probably new graphics cards/IP for Nintendo and MS next gen console, and the graphics for the Windows 7 phone based off the Xbox360 GPU. Wow, Let all that sink in for awhile. Assuming that this is what AMD is doing, and assuming that they actually pull all of this off, then this will probably be the biggest engineering feat/streak that the industry has ever known. Has ATI bitten off more than they can chew? And if not, how the heck is this the same company that put out R600 only a few years ago?

Also, if the ATI team can pull all this off, I think they're wasting a lot of energy. There is no real need to update Evergreen so quickly -- the family is more than competitive. They'd probably be better off splitting into two families/teams, one that only aims at gaming performance, and the other that only aims at stream performance. Fermi is a hybrid that fails at game performance (as an architecture in terms of die size and power consumption), but is almost definitely a better stream processor than the competition (at least in more complex scenarios where it's efficiency can beat out evergreens monster theoretical FLOPS). A family of processors aimed at only attaining high stream performance could do the same thing to Fermi in GPGPU that Evergreen is doing to Fermi in graphics. If you have the engineering resources for so many projects, you should try and elimate some of the overlap.

It depends, if they do release that hybrid, it might give them some time headroom to delay Northern Island a little while staying competitive as well as having new models for the holidays season and the associated revenues. Now, delaying Northern Island could have two advantages:

1. Let the 28nm process mature, run most respin if needed, and maybe make a move toward GF to further boost AMD revenues from their share in it.

2. Windows 8. Now this is a really far fetched rumor as MS rarely refresh their OS so fast, especially since Windows 7 is quite well accepted, but still if http://windows8news.com/2010/03/10/windows-8-to-be-release-july-2011-a-rumor/ proves to be true, then AMD is most likely already aware of it from its quite good relationship with Microsoft and could synchronize Northern Island release with the new Windows one.

Let push the rumor even further and assume that Windows 8 comes with DX12 and that MS is already working on XBox 3. Then maybe AMD wants to repeat the Evergreen's feat with DX11. I read an article just after DX11 got released about one feature (I cannot remember which) that was supposed to make it in 11 but didn't because of time constraints. However, the article stated that the Evergreen family already supported it. Assuming that feature gets included in DX12 (likely), then ATI already has a head-start, so maybe a DX12 Northern Island architecture that would also get used in the next XBox is a possibility. It would be a decent management move too imho.

 

[Daedalus685]

Absolutely



The rumors for SI/NI sound kinda promising, I'm digging the modular design bit (kind of more akin to CPU manufacturing, 2-6 big core modules linked by shared caches/uncores/stuff like that) but who knows if the 2x800 part rumored will be enough to combat a fermi refresh. I doubt it will do better than match the current fermi, and NV I suspect is working on a refresh... a 725mHx 512sp B1 Fermi would be a wonder to behold, and could possibly keep thermals manageable and not cost too much either. ATI'd have a tough battle fighting that one


Any rumors around regarding NVs plans? (fermi refresh, next gen, 28nm)

I don't think Fermi will be able to get the power consumption/heat lower without a shrink. Not in any meaningful way... They could perhaps tweak with a B1 to get a full 512 out at the current clock rates.. but any improvements to power and thermals would be cancelled out by the larger core count. It would still be huge after all.

No way to be sure how a new Fermi and a 6750 would stack up though..

 

[Kuzi]

I don't believe this Hybrid chip talk. If ATI does release a new card this fall on 40nm, it'll probably be a tweaked Cypress GPU without the cut backs on the original design. Which according to Anand's excellent "RV870 Story", would have been around 400mm^2 in size. This way ATI makes sure they won't face any delays, and a 400mm^2 GPU should not be a problem to produce 5-7 months from now on a mature 40nm TSMC process.

I feel like all ATI needs right now is a faster clocked 5870, say a 1GHz core and 2GB memory @1.4GHz (5.6GHz) GDDR5. Such a card would be about even in performance with the GTX480.

 

[RussianSensation]

there there, Scott

Wreckage is just disappointed he isn't going to be able to cook food on his GTX480...and he ordered it for its multi-purpose functionality:
http://www.legitreviews.com/article/1264/1/

I think he is just angry that ATI was able to launch all of their 5xxx series today with the culmination of Eyefinity 6 edition card, but NV doesn't have a single GT4xx series card out, not to mention months to wait for anything <$350.

 

[MrK6]

I'm not sure this news is a bad thing. Given how much TSMC botched the 5xxx series launch, I thought it was overly optimistic (or naive) of AMD to trust them with the new NI architecture on a new 28nm process. History shows that launching a new architecture on a new process isn't a good idea, with Fermi being the most recent example. However, that also might be the only way to actually launch said part (it was for Fermi, and maybe NI is that big/complicated too). Still, "Southern Islands" sounds neat, and it will be interesting to see what it brings to the table performancewise. Good find OP, thanks for posting .