AMD to manufacture console APU at GLF

[SiliconWars]

They are locked in a take-or-pay contract.

Which goes both ways - GF must provide a set amount of wafers for AMD's use every month.

This is going to be very useful for AMD from now on. It's not unreasonable to assume 20 million next-gen consoles will be sold in 2014. Looking at the other numbers...

Assume ~150 good die per wafer for the Xbox and PS4 combined, that's 133k wafers worth in 2014. With 40K wafers per month that's 3 whole months of production at GF, just on the consoles - and that's still without the China market. Even if the PC market collapses all AMD needs to do is sell 3/4rs at the same WSA they currently have and the fabs will be full. If things pick up on that front, or even more consoles are sold, they always have the opportunity to take more wafers.

TSMC couldn't give them that level of wafer allocation in 2014, not with Kabini and graphics there as well. 28nm is still big for TSMC's other players next year, really there is no way AMD could have provided enough consoles while being TSMC only. This is why the WSA is beneficial now. Try to imagine what Nvidia would have had to sacrifice had they been the console provider.

 

[mrmt]

Which goes both ways - GF must provide a set amount of wafers for AMD's use every month.

Is there? This GLF obligation was never mentioned in any Q&A or in whatever disclosures AMD made about the WSA. I would be glad if you could point out anything about that.

What we have in terms of evidences is that GLF does not have kind of obligation to AMD. Just look at the 28nm snafu. AMD had to reshuffle their entire line up, postponing steamroller and scrapping krishna and wichita, and GLF had to pay 0 for not providing the 28nm wafers on time.

Assume ~150 good die per wafer for the Xbox and PS4 combined, that's 133k wafers worth in 2014. With 40K wafers per month that's 3 whole months of production at GF, just on the consoles - and that's still without the China market. Even if the PC market collapses all AMD needs to do is sell 3/4rs at the same WSA they currently have and the fabs will be full. If things pick up on that front, or even more consoles are sold, they always have the opportunity to take more wafers.

Regardless of the math, it is clear that AMD would not be able to reach the WSA commitments with Richland and Kaveri only, and that they need something else to fill the quota. I have a feeling that it was just GLF inability to deliver a functional 28nm node that prevented the console APUs to be launched with GLF chips. Embedded, lagging edge products are exactly the kind of product suited for GLF.

Given that AMD "growth" opportunity (desktops) isn't really growing, I think this is a mostly urgent measure.

TSMC couldn't give them that level of wafer allocation in 2014, not with Kabini and graphics there as well. 28nm is still big for TSMC's other players next year, really there is no way AMD could have provided enough consoles while being TSMC only. This is why the WSA is beneficial now. Try to imagine what Nvidia would have had to sacrifice had they been the console provider.

They couldn't give the allocation, or couldn't AMD pay for it? I found impressive claims that TSMC would't be able to serve AMD, when they can serve Qualcomm, Mediatek, Nvidia and others. Some players bigger than AMD and TSMC total production makes AMD numbers look pale in comparison.

 

[SiliconWars]

What we have in terms of evidences is that GLF does not have kind of obligation to AMD. Just look at the 28nm snafu. AMD had to reshuffle their entire line up, postponing steamroller and scrapping krishna and wichita, and GLF had to pay 0 for not providing the 28nm wafers on time.

As I mentioned before, GF was shipping 28nm mobile products before AMD was shipping Kabini and Temash with TSMC, so clearly there has been errors made on both sides.

Regardless of the math, it is clear that AMD would not be able to reach the WSA commitments with Richland and Kaveri only

Clear how? Clear in that AMD has had no problems paying the 2013 WSA with Richland? There is nothing to suggest that AMD would not have been able to pay the 2014 WSA with Kaveri.

and that they need something else to fill the quota. I have a feeling that it was just GLF inability to deliver a functional 28nm node that prevented the console APUs to be launched with GLF chips. Embedded, lagging edge products are exactly the kind of product suited for GLF.

Large chips that don't require to be on the cutting edge are absolutely perfect for GF's 28nm, is what you mean. AMD will be getting ~30 extra die per wafer, that's $3K per wafer more on simple die size alone, before whatever else GF offers them. I'm sure the wafers themselves are much cheaper also, and that's why AMD is able to target raising margins from 15% to 20%.

Given that AMD "growth" opportunity (desktops) isn't really growing, I think this is a mostly urgent measure.

From your own link AMD clearly state that PC's will be down 10% in 2014, just like they said they'd be down in 2013 while Intel buried their heads in the sand and kept telling us that the back-half would rebound.

If they were worried about it they'd just have got a new WSA for $1 billion instead of $1.15 billion, don't you think? No, they were never worried about it because they are not struggling to pay the current one even without consoles.

They couldn't give the allocation, or couldn't AMD pay for it? I found impressive claims that TSMC would't be able to serve AMD, when they can serve Qualcomm, Mediatek, Nvidia and others. Some players bigger than AMD and TSMC total production makes AMD numbers look pale in comparison.

Yet up till recently TSMC was fully at capacity. It was only at the end of this year that things started to fall, with GF taking customers. Next year they'll need to ramp 20nm while 28nm is still the main node. It's asking a lot to expect AMD - who is already struggling to get enough wafers from TSMC - to get another 25k per month under that situation.

Clearly the consoles pay so there is no reason to believe that AMD "couldn't pay" for more. You're confused between can't pay and getting a much better deal, which is what they're getting at GF. I really don't understand why you do this to yourself mrmt - you realise that come Q2 2014 AMD will be showing a large increase in (console) margins making your predictions look hollow again? Or do you realise this? This is what is going to happen, fact.

That's not to say that something else won't go pear-shaped for AMD (it almost certainly will), but you're barking up the wrong tree with consoles at GF, big time.

 

[mrmt]

If they were worried about it they'd just have got a new WSA for $1 billion instead of $1.15 billion, don't you think? No, they were never worried about it because they are not struggling to pay the current one even without consoles.

So let me see... they order only 2/3 of their commitments in 3/4 of the year, but wait, that 1/3 they have to order is about Q1 demand, which is the weakest quarter for consumer semiconductors. I wonder how can you explain this funny math. Analysts couldn't, so they had to ask a lot of times how AMD was expecting to fulfill their WSA quota. I'm waiting for your explanation.

 

[SiliconWars]

So let me see... they order only 2/3 of their commitments in 3/4 of the year, but wait, that 1/3 they have to order is about Q1 demand, which is the weakest quarter for consumer semiconductors. I wonder how can you explain this funny math. Analysts couldn't, so they had to ask a lot of times how AMD was expecting to fulfill their WSA quota. I'm waiting for your explanation.

Well, they'll fulfil their WSA by buying wafers same as usual. I'm not really seeing the problem. Obviously they bought less wafers than normal this Q3 because Richland is on the way out, but that'll be corrected in Q4 with Kaveri coming in.

 

[sniffin]

[redacted]

If you have nothing nice to say...
-ViRGE

 

[Headfoot]

Good point!
* Worst deal I've ever seen. IIRC, contracts have to be mutually beneficial to be legal. Since the WSA remains in place it must be mutually beneficial somehow (at least on paper), but I don't see it.

Contracts don't have to mutually beneficial to be legal. You might be thinking of consideration, there just has to be consideration which can be as low as an exchange of $1. A contract can eventually become very detrimental to someone and still enforceable. That is in fact how things like futures, options and derivatives markets function fundamentally

 

[pablo87]

Kumar knows who butters his bread. The only executive who ever questioned GLF and the WSA was Dirk and he was fired shortly thereafter. Coincidence? I think not.

As far as the financial obligations are concerned, you only need to look at AMD's inventory to figure out where the WSA $1B+ commitment ended up in 2013.

Oh and by the way, in the pecking order of customers at TSMC, would AMD be a) first b) middle of the pack c) dead last by a wide margin? In other words, its a self-fulfilling prophecy that TSMC isn't supporting AMD...

AMD is uninvestable until private equity comes in and cleans house, starting with the Board of Directors. If the market corrects, this could very well happen in 2014. Or sooner if greed sets in for those who made a killing on Micron (they'll be looking for their next opportunity), or those who missed out. Let's hope so.

 

[pablo87]

Contracts don't have to mutually beneficial to be legal. You might be thinking of consideration, there just has to be consideration which can be as low as an exchange of $1. A contract can eventually become very detrimental to someone and still enforceable. That is in fact how things like futures, options and derivatives markets function fundamentally

Those are arms' length transactions. Arms' length transactions do not preclude bad deals by one party but its strictly a matter of competence (or lack thereof).

Whereas with AMD and GLF, almost everyone had a foot in 2 boats.

 

[Ajay]

AMD is uninvestable until private equity comes in and cleans house, starting with the Board of Directors. If the market corrects, this could very well happen in 2014. Or sooner if greed sets in for those who made a killing on Micron (they'll be looking for their next opportunity), or those who missed out. Let's hope so.

^^This, I hope. But does AMD's market cap need to shrink further for this to happen? Is this why you said 'if the market corrects'?. I've seen private equity do some amazing things with tech firms. AMD has some solid engineering and design teams as well as some key technical managers, but upper management and the BOD seem to be failing.

There are some really tough challenges for AMD. One is the WSA - is there any way to get out of that? And, second, I think AMD needs to be an IDM if it want's to succeed in the CPU business (they need to control the process tech that best fits the needs of the CPU division) - that would take a lot of cash. The ODM structure works fine for ASICs, but CPUs are different kind of beast.

 

[mrmt]

Kumar knows who butters his bread. The only executive who ever questioned GLF and the WSA was Dirk and he was fired shortly thereafter. Coincidence? I think not.

I don't think that questioning WSA was what caused of Dirk demise. Backing up derpdozer should. Dirk had a golden chance in his hands when in 2009 he killed the 45nm derpdozer. He could have killed the entire pipeline there and refocused AMD on Kabini and Stars. That he picked derpdozer and pushed the things to the last consequences.

 

[NTMBK]

I don't think that questioning WSA was what caused of Dirk demise. Backing up derpdozer should. Dirk had a golden chance in his hands when in 2009 he killed the 45nm derpdozer. He could have killed the entire pipeline there and refocused AMD on Kabini and Stars. That he picked derpdozer and pushed the things to the last consequences.

AMD did make a shrunk and improved Stars APU on 32nm. It was called Llano, and it performed significantly worse than Trinity. Piledriver was faster and more efficient than the Stars core it replaced on the same process node, contrary to your constant FUD campaign.

 

[Headfoot]

Those are arms' length transactions. Arms' length transactions do not preclude bad deals by one party but its strictly a matter of competence (or lack thereof).

Whereas with AMD and GLF, almost everyone had a foot in 2 boats.

You'd have a really hard time arguing that a contract between two highly sophisticated parties is a contract of adhesion, and that doesn't even carry any legal implications. Its just a scary word. Now if you're saying there's some bad faith, that's a separate matter and similarly very difficult to prove

 

[Ajay]

AMD did make a shrunk and improved Stars APU on 32nm. It was called Llano, and it performed significantly worse than Trinity. Piledriver was faster and more efficient than the Stars core it replaced on the same process node, contrary to your constant FUD campaign.

And how many man hours were put into the design of Llano compared to Bulldozer? Stars could have gotten a massive architectural overhaul for the $$s spent on Bulldozer (including the scrapped 45nm design). AMD could have added SMT for starters, for better server performance @ a low xtor penalty. We'll never know now, since AMD took a gamble on a 'blue sky' design and came up short.

 

[NTMBK]

And how many man hours were put into the design of Llano compared to Bulldozer? Stars could have gotten a massive architectural overhaul for the $$s spent on Bulldozer (including the scrapped 45nm design). AMD could have added SMT for starters, for better server performance @ a low xtor penalty. We'll never know now, since AMD took a gamble on a 'blue sky' design and came up short.

Potentially, yes. But we'll never know. Hey, what would have happened if Intel had tried to improve IPC instead of frequency on the follow up to Northwood? Maybe we'd still have a Pentium 4 based Intel architecture today.

At the time that the 45nm Bulldozer got canned, AMD had to make their decision based on the situation that they were in, and they felt that a 32nm Bulldozer would be better than a 32nm Stars core. Seems that they were vindicated. But I guess the real test will be in how the follow up architectures perform- whether the CMT concept has room to grow.

 

[mrmt]

AMD did make a shrunk and improved Stars APU on 32nm. It was called Llano, and it performed significantly worse than Trinity. Piledriver was faster and more efficient than the Stars core it replaced on the same process node, contrary to your constant FUD campaign.

And since when beating an ancient core like K10 is any kind of merit? In fact, you are attesting the size of the Bulldozer failure: an ancient core shrunk to 32nm is competitive against the might bulldozer. And you can bet that stars costs was just a fraction of bulldozer costs.

No wonder Dirk was fired.

 

[SiliconWars]

As far as the financial obligations are concerned, you only need to look at AMD's inventory to figure out where the WSA $1B+ commitment ended up in 2013.

AMD's historical inventories (m) -

Q3 2013 - WiP 667, FG 225, Total $922
Q2 2013 - WiP 516, FG 166, Total $711
Q1 2013 - WiP 422, FG 159, Total $613
Q4 2012 - WiP 547, FG 158, Total $744
Q3 2012 - WiP 557, FG 246, Total $833
Q2 2012 - WiP 411, FG 148, Total $585

Considering the consoles and Hawaii have just launched I think it's fair to say that your inventory explanation just doesn't add up. Finished goods are even less than in Q3 last year and the increase in Q3's work in progress will mostly be due to consoles and graphics, like they said.

Oh and by the way, in the pecking order of customers at TSMC, would AMD be a) first b) middle of the pack c) dead last by a wide margin? In other words, its a self-fulfilling prophecy that TSMC isn't supporting AMD...

What do you mean "TSMC isn't supporting AMD..."? AMD is a large and important customer for TSMC, probably contributing 5-10% of their entire revenue, not to mention historically being first to a new node and contributing R&D to that.

 

[NTMBK]

And since when beating an ancient core like K10 is any kind of merit? In fact, you are attesting the size of the Bulldozer failure: an ancient core shrunk to 32nm is competitive against the might bulldozer. And you can bet that stars costs was just a fraction of bulldozer costs.

No wonder Dirk was fired.

That... makes less than no sense.

Your initial argument: Instead of a 32nm Bulldozer, they should have focused on the Stars core.

My rebuttal: They did also develop Stars further, and it was significantly slower than Trinity on the same process.

Your response: Aha, performing better than a shrunken and improved Stars is no achievement at all!

What the heck are you actually trying to argue here?

 

[mrmt]

That... makes less than no sense.

Your initial argument: Instead of a 32nm Bulldozer, they should have focused on the Stars core.

My rebuttal: They did also develop Stars further, and it was significantly slower than Trinity on the same process.

Your response: Aha, performing better than a shrunken and improved Stars is no achievement at all!

What the heck are you actually trying to argue here?

That they could have spent money elsewhere. That investing in k10 would yield a far better chip or at least something far cheaper than bulldozer and still have comparable performance. 7 years of R&D just to get comparable performance of an ancient core, how can you defend that decision?

 

[SiliconWars]

That they could have spent money elsewhere. That investing in k10 would yield a far better chip or at least something far cheaper than bulldozer and still have comparable performance. 7 years of R&D just to get comparable performance of an ancient core, how can you defend that decision?

Piledriver is quite often 50% faster than Thuban (8350 vs 1100T) in MT loads, which is what the chip was designed for. That's what you should be comparing it to, not Bulldozer which was clearly flawed.

 

[NTMBK]

That they could have spent money elsewhere. That investing in k10 would yield a far better chip or at least something far cheaper than bulldozer and still have comparable performance. 7 years of R&D just to get comparable performance of an ancient core, how can you defend that decision?

If they had had that information on their desks 7 years previous then no, I couldn't justify that decision. But of course they didn't. They were trying to predict both their own hardware development over the next ~7 years and developments in software over the next ~10 years. They made their predictions and went with what it showed, and got to where they got to- which while better than Stars, yes, wasn't massively better.

But we're not talking about 7 years earlier. We're talking about when it became clear that the 45nm Bulldozer wasn't going to make the grade, and was going to take a lot longer in the oven. AMD's choice was to do one of the following:

a) Follow through with the Bulldozer research, launch the chip which was predicted to initially do significantly better than an improved Stars core on the same process node and then build on that

b) Can the entire project, go back to the drawing board and try to dig up some new way of getting performance out of the Stars core- putting the whole plan back and delaying their next generation performance core until goodness knows when, and potentially not even making any significant progress over the Husky core in Llano.

Your plan, Plan B, would have left AMD with only the Husky core from Llano for a significant amount of time as their "high performance" core. How do you think that alone would have done against Ivy Bridge?

 

[mrmt]

Piledriver is quite often 50% faster than Thuban (8350 vs 1100T) in MT loads, which is what the chip was designed for. That's what you should be comparing it to, not Bulldozer which was clearly flawed.

Quite often? Could you expand on the subject?

 

[NTMBK]

Piledriver is quite often 50% faster than Thuban (8350 vs 1100T) in MT loads, which is what the chip was designed for. That's what you should be comparing it to, not Bulldozer which was clearly flawed.

In all fairness, that's a bad comparison. Piledriver and Thuban were on different process nodes. Llano and Trinity is the closest comparison we have- same market segment, same process node.

 

[mrmt]

a) Follow through with the Bulldozer research, launch the chip which was predicted to initially do significantly better than an improved Stars core on the same process node and then build on that

b) Can the entire project, go back to the drawing board and try to dig up some new way of getting performance out of the Stars core- putting the whole plan back and delaying their next generation performance core until goodness knows when, and potentially not even making any significant progress over the Husky core in Llano.

Your plan, Plan B, would have left AMD with only the Husky core from Llano for a significant amount of time as their "high performance" core. How do you think that alone would have done against Ivy Bridge?

Bad, but Bulldozer isn't a significant improvement either, is it? And killing Bulldozer in 2009 would mean that AMD would have a brand new chip by 2013, and depending on the scope of this new chip that would have left the company with a far better cash position, especially if they started to make the SG&A cuts at that time. Instead AMD went full speed ahead with the R&D, didn't adjust the cost structure and launched Bulldozer.

Dirk was staring at a failure in 2009, and instead of canning it and move on he proceeded ahead. This is the decision that killed his job, not speaking against the WSA. WSA is killing AMD because the CPU business is becoming a shadow of its former self. If they weren't to pay charges, the WSA would be just bad.

 

[SiliconWars]

Sure but that's also unfair on BD/PD as the traditional architectural reasons for BD/PD was more cores. Trinity is nothing but a hack together of a good iGP and half a server chip. If I recall correctly, the BD architecture doesn't start coming into it's own until the 5th core and onwards.