TSMC confirms 40nm yield issues

[Idontcare]

Originally posted by: Wreckage

Originally posted by: Asianman
mod edit (non-"back on topic" part removed)

Back on topic, would the yield issues affect ATI's profit margins or TSMC's. Who actually takes the losses for the bad chips?

I'm sure TSMC can raise their prices to compensate. Especially if UMC is no where near 40nm.

The contract prices are locked, TSMC can't effectively raise prices in a dynamic fashion as some folks might envision it to be from the way you word that statement.

But the answer to the question of "who pays for poor yield" is that "we do"...the end customer.

How those dollars travel up-stream thru the retailer to the distributor to the designer and the fabber is a matter of contract.

TSMC sells wafers. By that I mean when you contract with TSMC you don't buy chips, you don't buy yield, you buy produced/finished wafers. Those wafers may have 20% yield, they may have 90% yield.

Now in the contract there are some (naturally) boundary conditions regarding how crappy a wafer or lot can be before it is scrapped inline instead of being considered sellable product from TSMC's point of view. And therein lies the rub when it comes to wafer pricing and yields.

A customer (ATI for example) might be willing to contractually accept more variability in yield (and a lower acceptable minimum yield number) in exchange for a lower price/wafer contract. For example let's say they accept any wafers with 20% yield or more in exchange for a price of $6k/wafer.

Now yield is a partly function of die-size (defect density, D0, related) so a 20% yielding wafer for ATI's 137mm^2 is a very different situation versus trying to yield 20% with a 500mm^2 die. So NV might have a contract that stipulates the minimum acceptable yield is 20% as well, but TSMC requires $7k/wafer for those 500mm^2 die. In effect TSMC is charging NV more (in my hypothetical example) than they are charging ATI to compensate for the yield discrepancy between the two chip designers.

But in the end who really pays for those wafers, be they $7k/wafer or $3k/wafer? We do. And that is why it is perfectly reasonable for us enthusiasts to talk about, and care about, die-size and TSMC's yields and HOL (health of the line).

 

[Wreckage]

Originally posted by: Idontcare


But the answer to the question of "who pays for poor yield" is that "we do"...the end customer.

Not necessarily, because the 40nm product still has to compete against a 45/50/55nm part. So a company may have to eat the extra cost of moving to a new process in order to keep competitive price wise.

I imagine TSMC covers much of the cost as part of R&D.

Either way the price we pay for a card is far more complicated than how much a wafer costs.

 

[Idontcare]

Originally posted by: ViRGE

Originally posted by: Idontcare

Originally posted by: thilan29

Originally posted by: Idontcare
UMC was relevant for GPU foundry pre-65nm...but they have fallen behind TSMC so badly when it comes to timeline of releasing leading-edge process technology that both ATI and NVidia have used TSMC exclusively for their GPU's since 65nm.

Thanks for the input. I knew they were in the GPU game before but didn't know what happened to them.

Yeah, to humor yourself google "UMC 45nm" and try and find a google hit dated anytime post June 2007.

I can't find anything dated 2008 let alone 2009. That's how bad it is at UMC at the moment.

Like Via or Ali based mobos, one year they are all the rage, two years later and you realize you haven't seen or heard anything about them in literally years.

How's Chartered doing these days?

Chartered is as equally behind TSMC as UMC. Chartered has a significant advantage though in that they are part of the IBM foundry ecosystem (a distinctly separate but comparable group of IDM's as to the IBM fab club) so they no further behind TSMC than is IBM. UMC is getting "help" as they are partnered directly with the remnants of TI's process development team from a couple years ago, but that relationship is much less geared towards the common goal that IBM's foundry ecosystem is geared so I would not be surprised if Chartered fields a viable 45nm foundry process before UMC.

SMIC was actually treading water pretty well too up to the 90nm node, then they just kind of fell off the roadmap for leading edge offerings for all the cost vs. volume reasons that UMC and chartered did at sub-65nm.

I'm really looking forward to seeing what comes out of the 32/28nm cycle foundry-wise. The debut of GlobalFoundries bulk-Si 32nm and the intersection of Intel's 45nm/32nm process tech into the GPU world can't help but shake things up markedly in the high-performance fabless IC industry.

 

[Adul]

I wonder if AMD will have their graphics chips fab at global foundries some day.

 

[Idontcare]

Originally posted by: Adul
I wonder if AMD will have their graphics chips fab at global foundries some day.

It would speak volumes about GlobaFoundries if they didn't.

 

[taltamir]

Originally posted by: Idontcare

Originally posted by: Wreckage

Originally posted by: Asianman
mod edit (non-"back on topic" part removed)

Back on topic, would the yield issues affect ATI's profit margins or TSMC's. Who actually takes the losses for the bad chips?

I'm sure TSMC can raise their prices to compensate. Especially if UMC is no where near 40nm.

The contract prices are locked, TSMC can't effectively raise prices in a dynamic fashion as some folks might envision it to be from the way you word that statement.

But it would be poor contract work indeed if they did not account for such a scenario. Especially considering what industry that is.

 

[Idontcare]

Originally posted by: taltamir

Originally posted by: Idontcare

Originally posted by: Wreckage

Originally posted by: Asianman
mod edit (non-"back on topic" part removed)

Back on topic, would the yield issues affect ATI's profit margins or TSMC's. Who actually takes the losses for the bad chips?

I'm sure TSMC can raise their prices to compensate. Especially if UMC is no where near 40nm.

The contract prices are locked, TSMC can't effectively raise prices in a dynamic fashion as some folks might envision it to be from the way you word that statement.

But it would be poor contract work indeed if they did not account for such a scenario. Especially considering what industry that is.

The contracts aren't engineered to be punitive in due course of their existence (they contain clauses regarding punitive actions when intentional breaches are made of course), its a working understanding that no one makes money if the chips aren't flowing from the fab at a pricepoint that enables devices which the end-user is willing to pay for.

That said, without competition TSMC does get to decide when in time and for how much money they are willing to bother to get those chips flowing from the fab. They aren't about to price themselves out of the opportunity shipping 40nm to more customers, at the same time they've got four years of R&D to start recouping plus the capex depreciation for the fab equipment.

It's a balancing act and to do it effectively requires the participants to be willing to take risks with each other's resolve to carry-thru on their commitments (commitments to ramp, commitments to place orders, etc).

The commodity memory segment of this industry is one area in which contracts mean very little, they get broken all the time whenever it benefits one party over the other and very very rarely do the broken contracts get pursued in court for enforcement.

But in the foundry world where business relationships make the difference between getting all your wafers or getting put on notice that you'll get zero wafers when capacity shortages happen (and they do happen) it really never pays to screw your foundry partner unless you are really ready to burn that bridge and never cross it again.

 

[taltamir]

punitive? what is punitive by accounting for the possibility of price cuts and competition?

 

[Idontcare]

Originally posted by: taltamir
punitive? what is punitive by accounting for the possibility of price cuts and competition?

It sounds like you are talking about contract negotiations then, not renegotiation's as I originally interpreted your post to imply.

The "accounting for pricecuts and competition" comes in when the fabless company is negotiating their wafer volumes and prices for the contract under negotiation.

Contracts get negotiated every quarter, sometimes less frequently but quarterly is pretty typical. If price cuts and competition are going to put pressure on the foundry contract it will only come to bear during the negotiation phase for the next contract, it has little bearing on the existing one.

Under very special extenuating circumstances contracts do get renegotiated during the course of very severe downturns, but again this is where bridges get burned and you very rarely see a fabless company take on such drastic measures as these foundries make sure you reap what you sow.

I assumed you were implying that existing contracts are worded to allow leverage for wiggle on existing pricing (or were bad contracts if they lack such wording) as you were responding to my post where I said the contracts are locked, which they are.

 

[taltamir]

i am saying that it does not NEED to require a renegotiation. It makes no sense for the company NOT to account for dynamic price changes in every contract, instead of renegotiating it quarterly, it could have a variety of built in conditions and their results... such as "if competitor releases product Z within X months than price is adjusted by Y%" and so on.

 

[Keysplayr]

Hmm. It seems you guys are more fascinated by the corporations, their contracts and money, and how they work, than the graphics cards themselves. I wonder if maybe there should be a separate forum created for corporate business discussions. Seems popular.

 

[Idontcare]

Originally posted by: taltamir
i am saying that it does not NEED to require a renegotiation. It makes no sense for the company NOT to account for dynamic price changes in every contract, instead of renegotiating it quarterly, it could have a variety of built in conditions and their results... such as "if competitor releases product Z within X months than price is adjusted by Y%" and so on.

Taltamir I understand what you are saying, hypothesizing really, but I am merely telling you how it is in the real world for this particular industry segment. I worked with TSMC, and the other foundries, as well as the fact TI (my employer at the time) was a foundry itself for SUN. The contracts in the industry evolved to be this way for more reasons than I could lay out in simple 3-4 paragraph forum style posting but they will generally favor the foundry when it comes to the complexity and what they deliver.

I'll just say that what you argue for really only benefits one side of the two-party contract, specifically the fabless design house side, and does not benefit the foundry outright. Contracts will generally be to the benefit of the foundry as they are the ones taking on the high expenditure risk (capex)...not looking to justify this statement here with my briefly worded post (no laughing! :laugh, merely am telling you this is how it is for these guys.

When TSMC gets some competition, which GF may provide, then the contract side may open up but I really doubt it. The fabless guys have to design their chips for a specific foundry's node starting about 2-3 yrs in advance. Once they (the fabless guy) are that invested into a particular foundry's process specs they can't just up and walk away from the table. On the flipside a foundry wants to keep their fabless accounts from taking their future (2-3 yrs away) IC designs to the foundry next door, so they too don't pull any hijinks's during contract negotiations (the 3 month ones) where cost suddenly spikes just because yield isn't coming in where it needs to be.

These are business partners in the fabless/foundry world, the foundry carries the capex risk and they get to generally set the tone of the contract and if the tone is going to favor anyone it will favor the foundry not the fabless.

Originally posted by: Keysplayr
Hmm. It seems you guys are more fascinated by the corporations, their contracts and money, and how they work, than the graphics cards themselves. I wonder if maybe there should be a separate forum created for corporate business discussions. Seems popular.

Since it is related to cost-structure and thusly a contributing factor to performance/cost metrics (which is what we do end up paying for in the end) personally I see it as no different than posts which discuss process node characteristics (die-size, clockspeed, yields, node timelines, etc) as it all pertains to the performance/cost of the underlying CPU, GPU, SSD, or ram, etc.

It would be interesting to see where a separate sub-forum on the topic would evolve, but at the same time it seems like it would be near impossible to extract and remove business strategy and policy discussion from any GPU thread relating to pricing and timeline, they are just too intrinsically intertwined IMO.

 

[Keysplayr]

I disagree. Discussing the production costs, contract negotiations/renegotiations, market share, or any cost to the corporations just seems to me that it is really far and away a different type of topic than meremy discussing graphics cards and their respective price performance. The most frequently asked question in this forum is "What card should I get?", not, "Which stock option should I choose. Just the way I see it. 40nm shouldn,t really concern the end user any further than heat, power consumption, overclockability. Cost is a factor of course, but not what it costs the corporations, just what it costs the end user.

 

[Pantalaimon]

Cost is a factor of course, but not what it costs the corporations, just what it costs the end user

One can affect the other though.

 

[Creig]

Originally posted by: Keysplayr
I disagree. Discussing the production costs, contract negotiations/renegotiations, market share, or any cost to the corporations just seems to me that it is really far and away a different type of topic than meremy discussing graphics cards and their respective price performance. The most frequently asked question in this forum is "What card should I get?", not, "Which stock option should I choose. Just the way I see it. 40nm shouldn,t really concern the end user any further than heat, power consumption, overclockability. Cost is a factor of course, but not what it costs the corporations, just what it costs the end user.

The costs incurred to design, engineer and produce a GPU inevitably get passed down to the end user, therefore it is a quite valid topic for Video. Obviously there is enough interest in the "behind the scenes" mechanics to keep a discussion on the topic going strong. I don't see any problem with it.

 

[Idontcare]

Originally posted by: Creig

Originally posted by: Keysplayr
I disagree. Discussing the production costs, contract negotiations/renegotiations, market share, or any cost to the corporations just seems to me that it is really far and away a different type of topic than meremy discussing graphics cards and their respective price performance. The most frequently asked question in this forum is "What card should I get?", not, "Which stock option should I choose. Just the way I see it. 40nm shouldn,t really concern the end user any further than heat, power consumption, overclockability. Cost is a factor of course, but not what it costs the corporations, just what it costs the end user.

The costs incurred to design, engineer and produce a GPU inevitably get passed down to the end user, therefore it is a quite valid topic for Video. Obviously there is enough interest in the "behind the scenes" mechanics to keep a discussion on the topic going strong. I don't see any problem with it.

TBH I'm kinda baffled why anyone would want to stifle or imply discouragement towards the discussion of the topic of business practice, etc.

Usually enthusiasts thirst for access to more knowledge about the industry, it's why I come here. To learn from others and, where applicable, to share what I know from my own firsthand experiences in life.

I can't tell if keys is carefully wording his posts with his hand on the moderator hat, about to switch over and say "keep this on-topic, and business matters are not on topic IMO"...it definitely gives me that feel though, so I be a bit befuddled at the moment as to where this is headed.

 

[Keysplayr]

Guys, don't read into this that much. I just wanted to get a feel for the need to request another forum or not bother. Of course I do not wish to stifle these discussions. Yes, these types of discussions do often serve to derail threads, hence my thoughts on a separate forum. Just an idea. Remove all of your tin foil hats guys

 

[Forumpanda]

Originally posted by: Keysplayr 40nm shouldn,t really concern the end user any further than heat, power consumption, overclockability.

But then you can also argue that this is not really a forum for the average end user.
I can do math and understand numbers myself, I don't need someone to break down the performance/cost or performance/watt for me.

I come here solely to read these kinds of threads, because I like to see where the industry is heading and what kind of road bumps it is hitting.

Now if only i could find someone to clone IDC so I could get more of his posts (No offense to anyone else ), this forum would be the best for what I am here for.

edit: posted before I saw your last post, just saying, these kinds of discussions happen because this is the kind of stuff that is not on every forum out there.

 

[Idontcare]

Thanks for clarifying keys, thankfully no tinfoil was sacrificed in the creation of the side-topic of creating another sub-forum...FWIW, and thilan please forgive the continuance of this side-topic we'll definitely take it to PM if you prefer, creating a sub-forum relating to business practice would not be a bad thing IMO, I'm just saying I don't think creating one would necessarily preclude the topic itself from invariably and legitimately cropping up in threads in the specific sub-forums where vendor names and their release timelines are discussed.

In CPU forum for example we have a fairly consistent re-emergence of the topic of electromigration and lifetime versus temps, etc basically every time a new node is released. It could easily be argued that such discussions belong in the Highly Technical sub-forum, or perhaps a dedicated Process Technology sub-forum needs to be created, but no one really complains about having an EM thread in the CPU forum as lots of folks seem to like to be exposed to the dialogue even if in passive reading manner.

Nevertheless this is the Video forum, not the CPU forum, and if process tech, yields, cost structures discussions are not really what the readership wants to be exposed to here then that is feedback I certainly need to hear (and I imagine a few others with technically-geared backgrounds would like to know too) as the last thing I want to be doing is be wasting my time creating posts that just serve to waste the readers time as well if they aren't interested in reading about that stuff. (I'm serious about this, maybe one of the video mods could make a poll for the forum and let the people vote?)

 

[thilanliyan]

Originally posted by: Idontcare
and thilan please forgive the continuance of this side-topic we'll definitely take it to PM if you prefer,

No worries...feel free to continue (plus I do not think it's my decision to make whether you guys can continue or not ). Discussing some of the economics is better than some of the flame wars that we see around here. I'm learning new things as well.

 

[HurleyBird]

Originally posted by: AsianmanBack on topic, would the yield issues affect ATI's profit margins or TSMC's. Who actually takes the losses for the bad chips?

I'd guess that costs would be split. I mean, if TSMC covered all the yield costs than everyone would be able to make huge chips at their expense. If ATI/NVDIA took the hit for yields then TMSC would have little reason to improve it's yields, which would drive companies to other fabs. I could be completely wrong as well though, no one really knows the details of the contracts. Each contract might be completely different too.

 

[taltamir]

really far and away a different type of topic than meremy discussing graphics cards and their respective price performance

for CURRENT cards, absolutely, but when discussing the potential price and performance of FUTURE cards build on future tech, the business practices are about the only thing we CAN discuss in order to attempt to make predictions.

 

[Keysplayr]

Originally posted by: taltamir

really far and away a different type of topic than meremy discussing graphics cards and their respective price performance

for CURRENT cards, absolutely, but when discussing the potential price and performance of FUTURE cards build on future tech, the business practices are about the only thing we CAN discuss in order to attempt to make predictions.

No, you really can't. Because you don't know what anything truly costs anybody. Can you see the key ingredient missing for this discussion type? If you had some hard numbers to work with, then that would be great. But we don't. Unless somebody with some inside information at TSMC wishes to make an appearance.

 

[taltamir]

well, i said its the only thing you can discuss, not that its gonna be an accurate discussion .
Predicting the future is an inaccurate art, which is why I use only the highest grade crystal balls .

 

[lopri]

What is the next logical step for GPUs after 40nm? 32nm? 28nm?