EETimes: ST plans for Dresden FDSOI production

[Ajay]

So you say that the next server processors will be 28nm Piledriver cores? Not really believable if they already taped-out SR-cores for Kaveri, and Richland is still only a 32nm Piledriver design. However AMD has not stated anything explicit so far, so I'll give the 28nm-Piledriver-idea an outsider's chance of 10% ;-)

If AtenRa posted that then he made a typo. He knows what is expected for the next iteration of Bulldozer and that is Steamroller. Based on what we know (not a lot) it'll be 28nm.

I sure hope that it's FD-SOI, or it will have no chance of being competitive with Intel. If it is built on FD-SOI and the information on performance is correct, and the estimated performance advances of SR are correct, them SR will actually be a pretty darn good CPU (>2x the performance/watt of PD, unless clock rates fall @ 28nm, which appears to be what happened with 28nm bulk). That would be an amazing result (probably too amazing).

<total speculation>
I think that anything above +35%/watt would be good an would meet the 15% increase in power/performance per year that AMD set as a baseline. If Jim Keller's team and GloFo deliver then we could be looking at a CPU that would bring enthusiasts back to AMD (say, +50% performance (ST) and -25% on power compared to PD). I don't think that this is going to happen on the desktop because AMD would need two sets of masks - one for a 4 module desktop and another for >4 module server MPU). AMD doesn't have the cash or resources for that, ATM.
</total speculation>

 

[AtenRa]

So you say that the next server processors will be 28nm Piledriver cores? Not really believable if they already taped-out SR-cores for Kaveri, and Richland is still only a 32nm Piledriver design. However AMD has not stated anything explicit so far, so I'll give the 28nm-Piledriver-idea an outsider's chance of 10% ;-)

I said, no way that SR is 28nm.

 

[Idontcare]

I sure hope that it's FD-SOI, or it will have no chance of being competitive with Intel. If it is built on FD-SOI and the information on performance is correct, and the estimated performance advances of SR are correct, them SR will actually be a pretty darn good CPU (>2x the performance/watt of PD, unless clock rates fall @ 28nm, which appears to be what happened with 28nm bulk). That would be an amazing result (probably too amazing).

When I look at my FX8350 box, literally sitting next to my 3770k box, I'm left with the conclusion that the FX8350 needs is a shrink and it would be so much more potent.

I'm not convinced taking 32nm SOI piledriver and shrinking it to 28nm bulk-Si is going to deliver the kind of reduction in power consumption that it needs to have in order to improve its positioning relative to 22nm Haswell.

If AMD is able to put out a piledriver shrink on 28nm FD-SOI then it could be quite the step forward for AMD.

I said, no way that SR is 28nm.

I must be totally confused as to what AMD is doing on 28nm if SR is not coming on 28nm. They already have the piledriver core for 32nm.

They would be flat-out wasting an opportunity to improve the microarchitecture if they chose to shrink piledriver to 28nm but did absolutely nothing to the core.

And if they did do something to the core then they are probably going to refer to that core as "Steamroller" given that steamroller is the only microarchitectural increment listed on all of AMD's roadmaps to date.

So how can steamroller not be a 28nm core? And if steamroller is not taped out for 28nm, then what is?

 

[Abwx]

The APU version of SR will be 28nm , this was displayed at CES
but true that nothing has been said about an eventual Vishera
replacement and server siblings.

 

[Idontcare]

The APU version of SR will be 28nm , this was displayed at CES
but true that nothing has been said about an eventual Vishera
replacement and server siblings.

That makes sense. But in that regard it would make no sense for AMD to expend resources developing steamroller for 28nm and then doubling the expense of shrinking the core for smaller nodes without altering/improving the core along the way.

In other words, the only thing that makes sense to me is that Piledriver is 32nm, Steamroller is 28nm, and Excavator (if it happens) is 20nm.

 

[Third_Eye]

Hmm, this could be huge for AMD. They can take existing 28nm designs and migrate them to 28nm FD-SOI w/30% higher performance or much lower power without using a large design team for each. AMD can reduce its node usage to only 2 because they can get the functionality of 4 nodes out of those two.

Still, nothing is certain since Dresden hasn't even started the conversion yet:

Nope. 0% chance of it happening.

AMD is now wisely backing off the SOI bandwagon. IBM was basically having a laugh with AMD doing the SOI hard talk as well as manufacturing a mass market product on SOI and trash talking bulk while it had a parallel bulk consortium consisting of Samsung, Renesas, Chartered Semi, Toshiba and STMicro.

STM has been promising too many things starting with Full Depletion. Right now the top 3 manufacturing processes are all Bulk HKMG

Intel 22nm Trigate Bulk
TSMC 28nm Planar Bulk
Upcoming Common Platform[IBM,SMSG,GF] 28nm Planar Bulk (Exynos Octa, SR )

Let STM commercialize the FD-SOI technology and release products on it.

Wanna know how I am 100% certain.
a) AMD's Feb 2012 Financial Analysts Day presentation
b) Q4 2012 earning conference calls
c) 2012 Dec Wafer Supply Agreement (WSA) which involved AMD paying 320M USD for not buying any wafers for Q4 2012.

Page 6.
<begin>
.. Separately, AMD will move to standard 28nm process technology and significantly reduce reimbursements to GF for future research and development costs.
&#8211;
We anticipate these savings will be approximately ~$20M per quarter during the next several years which also helps achieve our OPEX target of $450M by Q3 2013
..
<end>

If FD-SOI is so easy as STM suggests it is and yields pretty good, AMD can think of porting their Pile Driver Opterons [which are stuck at 32nm PD_SOI till end of 2014] to 28nm FD-SOI and then think of moving their more common products to that. Else it is a waste of money from AMD's perspective, which will accelerate its bleeding...

 

[SiliconWars]

That makes sense. But in that regard it would make no sense for AMD to expend resources developing steamroller for 28nm and then doubling the expense of shrinking the core for smaller nodes without altering/improving the core along the way.

In other words, the only thing that makes sense to me is that Piledriver is 32nm, Steamroller is 28nm, and Excavator (if it happens) is 20nm.

I agree with this. For steamroller to be 28nm for APU's but 32nm for cpu would be quite bizarre.

 

[Ajay]

Nope. 0% chance of it happening.

AMD is now wisely backing off the SOI bandwagon. IBM was basically having a laugh with AMD doing the SOI hard talk as well as manufacturing a mass market product on SOI and trash talking bulk while it had a parallel bulk consortium consisting of Samsung, Renesas, Chartered Semi, Toshiba and STMicro.

STM has been promising too many things starting with Full Depletion. Right now the top 3 manufacturing processes are all Bulk HKMG

Intel 22nm Trigate Bulk
TSMC 28nm Planar Bulk
Upcoming Common Platform[IBM,SMSG,GF] 28nm Planar Bulk (Exynos Octa, SR )

Let STM commercialize the FD-SOI technology and release products on it.

Wanna know how I am 100% certain.
a) AMD's Feb 2012 Financial Analysts Day presentation
b) Q4 2012 earning conference calls
c) 2012 Dec Wafer Supply Agreement (WSA) which involved AMD paying 320M USD for not buying any wafers for Q4 2012.

Page 6.
<begin>
.. Separately, AMD will move to standard 28nm process technology and significantly reduce reimbursements to GF for future research and development costs.
–
We anticipate these savings will be approximately ~$20M per quarter during the next several years which also helps achieve our OPEX target of $450M by Q3 2013
..
<end>

If FD-SOI is so easy as STM suggests it is and yields pretty good, AMD can think of porting their Pile Driver Opterons [which are stuck at 32nm PD_SOI till end of 2014] to 28nm FD-SOI and then think of moving their more common products to that. Else it is a waste of money from AMD's perspective, which will accelerate its bleeding...

So bizarre. If GF is just a foundry partner, why is AMD paying for GF's R&D. Ruiz must have been the worst negotiator ever - that or he got something else out of the deal.

 

[mrmt]

So bizarre. If GF is just a foundry partner, why is AMD paying for GF's R&D. Ruiz must have been the worst negotiator ever - that or he got something else out of the deal.

Debt. AMD offloaded a lot of debt to GLF. You cannot be tough with your banker.

 

[Ajay]

Debt. AMD offloaded a lot of debt to GLF. You cannot be tough with your banker.

Ah, I see. Too bad, if the claims for FD-SOI are roughly accurate, that would be a good thing for AMD, but paying an extra $80M/yr to develop it is a bad move, financially, for AMD right now

I'd love to see them shift their focus to mobile, but still crank out some competitive 'big core' products every two years, but it just can't happen with their eviscerated R&D budget.

 

[Idontcare]

So bizarre. If GF is just a foundry partner, why is AMD paying for GF's R&D. Ruiz must have been the worst negotiator ever - that or he got something else out of the deal.

Not the worst, the best. Remember Ruiz's plans were for him to leave AMD and be chairman of GloFo.

Sure he made life hard at AMD, but that was his successor's problem. He was making life a lot easier for himself at GloFo by setting the stage in advance

And yes, it is bizarre for a foundry to require a customer to specifically pay for R&D without the structure entailing some manner of "equity transfer" the likes of what ASML did with Intel, TSMC and Samsung.

I'll give you a real-world firsthand example. At TI we were a foundry for SUN Microsystems. We made three different variants of each node, the same as every other foundry. (low power mobile, high performance mobile, and high performance microprocessor)

Now at TI we only used the first two for our own products. The lower power mobile and high performance mobile. The high performance microprocessor sub-node was built solely and exclusively for SUN. And we never charged them for the R&D.

Oh sure they paid for it, but they paid for it the same as every customer pays for the R&D that went into the products they buy (it is amortized into the cost structure which sets the pricepoint for the purchase price). None of us pay Samsung an annual "R&D surcharge" above and beyond the purchase price of a Samsung LCD screen from Newegg. Such a business model is absurd.

And it is absurd in the foundry markets as well...accept when it comes to GloFo.

And it is to their own demise that they keep doing these atypical one-off weird contracts that only seemed designed to penalize the customer. That is exactly the opposite kind of public perception you want as a foundry that needs new customers.

Wafer supply agreements, take or pay contracts, R&D surcharges...absolutely none of this screams "we want customers, we prioritize customer roadmaps over our own, we know what we are doing and how to do business as a foundry".

Is it any wonder TSMC nailed 90% of 28nm tapeouts and has 100% 28nm marketshare for over a year now? GloFo is doing themselves absolutely no favors in pursuing the execution of the contract that AMD signed, even if they are in their rights to do so, because it is literally pushing customers into TSMC's waiting arms.

 

[mrmt]

I'd love to see them shift their focus to mobile, but still crank out some competitive 'big core' products every two years, but it just can't happen with their eviscerated R&D budget.

I don't see the their R&D budget as a problem per se for the big core products.

Regardless of how mauled are AMD R&D resources, there is still a lot of R&D happening there. Bulldozer is slipping on the priority list because AMD thinks it can get better returns focusing resources elsewhere (Kabini, Brazos, etc).

Bulldozer was supposed to be a high volume architecture, and yet one year after launch they couldn't ramp up production to 10 millions per year, while Brazos needs only two quarters to get to the same numbers. Also Kabini can get a small share of tablets, a fast growing market, while Bulldozer wouldn't fit on the BoM even if they had low enough power consumption.

To sum up, the reason why big core is fading on AMD because Kabini and ARM are better bang for the buck, or at least AMD management thinks so.

 

[Hitman928]

Is it any wonder TSMC nailed 90% of 28nm tapeouts and has 100% 28nm marketshare for over a year now? GloFo is doing themselves absolutely no favors in pursuing the execution of the contract that AMD signed, even if they are in their rights to do so, because it is literally pushing customers into TSMC's waiting arms.

And the legacy of failure and ruin that is H Ruiz grows with each passing year and each new company he touches. It's too bad Motorola and AMD couldn't get rid of him sooner than they did.

 

[Ajay]

Not the worst, the best. Remember Ruiz's plans were for him to leave AMD and be chairman of GloFo.

Is it any wonder TSMC nailed 90% of 28nm tapeouts and has 100% 28nm marketshare for over a year now? GloFo is doing themselves absolutely no favors in pursuing the execution of the contract that AMD signed, even if they are in their rights to do so, because it is literally pushing customers into TSMC's waiting arms.

Forgot about Ruiz' plans to run GloFo. And yeah, GF has been proudly announcing it relationship with ARM, it 28nm process, it's supposed fast tracking of their 14nm node - and yet I haven't read about any new customers. Qualcomm was talking to GF for 28nm production (since they couldn't get enough from TSMC), but I don't know if that will happen - I haven't read about any deal lately.

If AMD dies, GloFo with die with them.

 

[Ajay]

I don't see the their R&D budget as a problem per se for the big core products.

Regardless of how mauled are AMD R&D resources, there is still a lot of R&D happening there. Bulldozer is slipping on the priority list because AMD thinks it can get better returns focusing resources elsewhere (Kabini, Brazos, etc).

Bulldozer was supposed to be a high volume architecture, and yet one year after launch they couldn't ramp up production to 10 millions per year, while Brazos needs only two quarters to get to the same numbers. Also Kabini can get a small share of tablets, a fast growing market, while Bulldozer wouldn't fit on the BoM even if they had low enough power consumption.

To sum up, the reason why big core is fading on AMD because Kabini and ARM are better bang for the buck, or at least AMD management thinks so.

While what you said has the ring of truth to it, SR Opterons were supposed to have a substantially higher IPC than PD. As a plug and play replacement, I think it would have sold well. Seeing how PD boosted Desktop sales over BD, SR would likely have done even better - but I don't know what AMD got as customer feedback, so it's possible that the feedback was that there weren't going to be enough takers and Kabini and Temash were looking to have a much high ROI.

The big core work is ongoing, and AMD will have something new out in 2014; I imagine that the big core team has shrunken considerably given the time line, otherwise the engineers would be twiddling their thumbs.

 

[NostaSeronx]

AMD work orders for GlobalFoundries are only SLP and HPP on the 28-nm node on the bulk process only.

(I know I previously said it was LPH, sad fate.)

And yeah, GF has been proudly announcing it relationship with ARM, it 28nm process, it's supposed fast tracking of their 14nm node - and yet I haven't read about any new customers.

There are six other customers other than AMD, IBM, ARM, Qualcomm, ST Microelectronics. That can easily be googled and confirmed on their website via a Press Release. One actually compares their product against Ivy Bridge, I think or Sandy Bridge.

 

[GreenChile]

Let's not forget that developing a RnD technology concept is not the same as developing a high yielding fully manufacturable process technology. STMicro is trying to sell a technology concept which still needs a lot of work to make it cost effective and functional for HVM.

They say GF could be ready to begin production as soon as Q4 2013 if they start development now. This is a best case scenario. Real world readiness would likely be 2014 at the earliest.

Bottom line AMD could not benefit from this near term.

 

[Abwx]

AMD work orders for GlobalFoundries are only SLP and HPP on the 28-nm node on the bulk process only.

(I know I previously said it was LPH, sad fate.)There are six other customers other than AMD, IBM, ARM, Qualcomm, ST Microelectronics. That can easily be googled and confirmed on their website via a Press Release. One actually compares their product against Ivy Bridge, I think or Sandy Bridge.

So far , GF 28nm process has its fet transistors Spice parameters
that are not definitive so anyone trying to design a circuit using
this process will have a functionnal circuit but its simulations of
caracteristics at high frequency wont be as accurate as with a fixed
process.

That s what i did understand from a recent interview of a GF
process engineer.

No wonder that there s no announced product using their process ,
as if it was really functionnal it would have been used for Kabini.

I may be wrong but do not expect 28nm products before
late 2013 , probably just in time for Kaveri.

 

[mrmt]

Seeing how PD boosted Desktop sales over BD, SR would likely have done even better - but I don't know what AMD got as customer feedback, so it's possible that the feedback was that there weren't going to be enough takers and Kabini and Temash were looking to have a much high ROI.

Yes, it is clear that Steamroller would have been better than Piledriver now, but, did Piledriver boost sales by that much? AMD gross margins improved 1% from last quarter, meaning only 11 million in profits from this new margin. Also they are shrinking K10 shipments, and we don't know in what conditions they were selling this inventory, but probably was at very low prices and very low margins.

So while there is truth in the sentence "our situation in desktops improved", the fundamentals of this improvement aren't the game changing AMD needs, but most likely a natural improvement that doesn't necessarily reflect a new trend.

The big core work is ongoing, and AMD will have something new out in 2014; I imagine that the big core team has shrunken considerably given the time line, otherwise the engineers would be twiddling their thumbs.

AMD went from being a 14.000 people company developing three product lines (GPU, big core, small core) to a 10.000 people company developing five product lines (GPU, big core, small core, ARM, micro servers). They are stretched by a lot, and this does explain the heavy user of synthesis tools by current and future products.

 

[Ajay]

Yes, it is clear that Steamroller would have been better than Piledriver now, but, did Piledriver boost sales by that much? AMD gross margins improved 1% from last quarter, meaning only 11 million in profits from this new margin. Also they are shrinking K10 shipments, and we don't know in what conditions they were selling this inventory, but probably was at very low prices and very low margins.

So while there is truth in the sentence "our situation in desktops improved", the fundamentals of this improvement aren't the game changing AMD needs, but most likely a natural improvement that doesn't necessarily reflect a new trend.

AMD went from being a 14.000 people company developing three product lines (GPU, big core, small core) to a 10.000 people company developing five product lines (GPU, big core, small core, ARM, micro servers). They are stretched by a lot, and this does explain the heavy user of synthesis tools by current and future products.

Every time there is the possibility for a game changer at AMD, reality comes in and crushes it. They just can't seem to win for losing :'(

 

[ShintaiDK]

Every time there is the possibility for a game changer at AMD, reality comes in and crushes it. They just can't seem to win for losing :'(

They simply dont have the size or money for it. AMD seems only to be able to succeed when a miracle happens. Perfect execution, everythign flawless and on the same time the competition needs to do everything wrong. It happend one time, and its unlikely to ever happen again.

 

[mrmt]

Every time there is the possibility for a game changer at AMD, reality comes in and crushes it. They just can't seem to win for losing :'(

Well, the problem isn't with reality but with the expectations created for future products.

When we see this amount of hot air coming from a company and not living up to its expectations, it is clear that this company has a management problem. When you don't live up to your hype, it damages you credibility with everyone, from consumers, to OEMs, partners and corporate buyers.

But why does this happen? I see a two-pronged problem:

1) AMD marketing department is is hype-addicted. From send drones to the forums, hiring firms to post on forums, recruiting users to make good posts for AMD, fake benchmarks, the list is long, very long.

2) It is clear that there were a lot of internal disputes between engineering teams and that dispute spilled to the market team. The bulldozer engineering team had to override internal controls to make Bulldozer happen, and they did that using overly optimistic projections, and when a hype-addicted marketing department gets its hands on overly-optimistic projections the sky is the limit.

Rory did change the things in the marketing department. While we don't see the the amount of hot air we saw at Dirk's term, and he quickly killed the illusion that AMD would carry on competing with Intel, AMD marketing message is filled the marketing message with empty buzzwords.

What they are doing now with their product is focusing on the strong points of their products and cherry picking scenarios to give improvement estimates of their future products. While it does not always get a honest picture for their consumers, there isn't much they can do outside this when their products are at disadvantage in a lot of critical metrics. All in all the marketing message is better calibrated now.

What some AMD fans must accept is that the company isn't trying to pull out a miracle they are waiting, but is trying to develop new products to pay the bills, and those new products won't be enthusiasts products.

 

[AtenRa]

It happend one time, and its unlikely to ever happen again.

32nm Trinity vs 22nm IvyBridge

AMD A8-5600K is ~42% faster than Core i7 3770K consuming only 14,5% more power.

You are right in one thing, Intel really needs a miracle to catch AMDs APUs in Graphics performance. :whiste:

 

[Abwx]

But why does this happen? I see a two-pronged problem:

1) AMD marketing department is is hype-addicted. From send drones to the forums, hiring firms to post on forums, recruiting users to make good posts for AMD, fake benchmarks, the list is long, very long.

[redacted]

I'm just going to stop you here. You have 4 infractions in less than 2 weeks in this forum alone, all for repeatedly attacking other posters and thread crapping. Since you cannot or will not play nicely with others, please enjoy some quality vacation time
-ViRGE

 

[SocketF]

So bizarre. If GF is just a foundry partner, why is AMD paying for GF's R&D. Ruiz must have been the worst negotiator ever - that or he got something else out of the deal.

AMD wanted its own 28nm SHP process, i.e. PD-SOI. It was shown occasionally on some GF slides. Nobody else wanted to have that process, thus AMD obviously would have to pay extra for it.

Now AMD and GF agreed that AMD will use only "standard processes" of GF in the future. Some people misinterpret that as bulk, but it basically just means that they will use anything of GF's "standard"
offers.

If GF offers FD-SOI then they could use it.

@Third_Eye:

Right now the top 3 manufacturing processes are all Bulk HKMG

Intel 22nm Trigate Bulk
TSMC 28nm Planar Bulk
Upcoming Common Platform[IBM,SMSG,GF] 28nm Planar Bulk (Exynos Octa, SR )

Trigate is not easy to make. It has bad yields and is overall very expensive. Intel can use it because they have lots of $ and because they can sell their processors for $$$$, too.
It's not an option for AMD. AMD's CEO stated officially that AMD will try to reduce mask layers. Coincidentally, FD-SOI reduces the necessary mask-layers ;-)

Let STM commercialize the FD-SOI technology and release products on it.

Already here:

http://www.advancedsubstratenews.co...netablet-chip-at-vegas-a-great-start-to-2013/