EETimes: TSMC estimates they hold 90% of the world's pending 28nm tape-outs

[Idontcare]

TSMC cannot even cope with demand almost 3 years after introducing the worlds worst process with their 40nm bulk. There are numerous sources on this, google it.

The ramp rates on 40nm have hardly been an outlier compared to historic norms.

40nm has not been blockbuster, nor has it been dead on arrival.

They did have capacity ramp issues in the 3rd quarter as AMD made a lot of justified publicity regarding. But I haven't heard much being said about it since.

You can call it "the worlds worst process" but the argument here really is hollow as TSMC would not have been capacity constrained if no one wanted it. If it was really as bad as you want to claim it was/is then why are so many customers using it? And why are so many customers lining up to use their 28nm process?

TSMC's 40nm can't be all that bad if even AMD was willing to produce 40nm bobcat with it. Surely AMD wouldn't make such an egregious error as to go with "the worlds worst process" for debuting their first Fusion product, would they?

 

[Idontcare]

As shown by an earlier chart, TSMC had 4 times the revenue of GF last year. GF was third by a slight margin behind UMC. For all intents and purposes UMC and GF had about the same revenue last year. Although UMC lost perhaps their biggest costumer for 28nnm in Xilinx to USMC (http://www.eetimes.com/electronics-news/4087890/Xilinx-confirms-Samsung-TSMC-in-UMC-out-at-28-nm) so they seem to be hurting for custers on their 28nm node as well.

I do really like their hybrid approach to HKMG though. They actually run a gate-first process, with HKMG material tailored to n-type gates, then remove the material at the p-type gates and replace it with p-type tailored HKMG material. (IIRC, I can't seem to open any of the articles about this from behind my firewall at the moment).

There was a good article on this (UMC's hybrid gate-first/gate-last integration) up on semiconductor.net before Reed pulled the site instead of archiving the articles.

Here's UMC's own PR on it: http://www.umc.com/english/news/2009/20091210.asp

Bah...nothing comes up in google.

If you've the time though I did find this pdf to be an entertaining read (in a good way!).

http://www.iwailab.ep.titech.ac.jp/pdf/iwaironbun/0809bogota.pdf

 

[jvroig]

You won't see IBM give up their fab and go fabless until they are assured and have confidence in their having access to the process tech they need to keep their server hardware development alive (<- that's not just my opinion, our CTO at TI at the time came from IBM and he told us this)

Agreed. I have heard these same sentiments from "the father of AS/400", Frank Soltis (in a very "geek-dream" moment, we even have a picture together, after I attended a symposium/talk years ago that he conducted).

Although he seemed to be greatly exaggerating their R&D advantage with mention of how their R&D is already playing around with below 1nm or something else that's equally ridiculous (far as I can recall; like I said, this was a couple of years ago or so, and that's Watson, you know how they are; as a sort of disclaimer, I like the idea of having something like the WRC exist, since R&D without the baggage of producing something to bring in $ is ok - it is as close to being "ideal" as possible, if the goal was advancement of tech for the sake of tech, instead of for the sake of dollars; but I digress...) , it was quite obvious that none of it will be realistic if they were to go fabless.

Given that no pure-play foundry will ever put up with the whims of WRC, find it highly unlikely that IBM would ever be satisfied becoming fabless like AMD.

Whether it will work out great for AMD remains to be seen in a span of years. I hope it doesn't turn out so bad. I currently have no indication to judge whether it will fail or not, so I can't even guess at this point. 3 years down the line, probably.

I'm still divided on whether I wish TSMC to retain a vast majority of the pie, or if I want GF (or other foundries) to catch up and eat a lot more of the pie. On one hand, the foundry business is expensive, and having one clear leader means that company has enough funds (naturally, this means customers riding on them) to keep on improving process tech on a more timely manner. On the other hand, a more equal distribution between two players might not be so bad, allowing just enough funds to accommodate advancements in process tech without affording the luxury of slumping around. I lean more on the first, since as it is, foundries already can't afford to slump around if they ever dream of being (or remaining) relevant.

 

[dma0991]

If you've the time though I did find this pdf to be an entertaining read (in a good way!).

http://www.iwailab.ep.titech.ac.jp/pdf/iwaironbun/0809bogota.pdf

From that I now know the possible limit is around 5-10 nm and most probably move on to other materials for future processors.

Not a very well made pdf but the amount of information that it contains is awesome. Thanks :thumbsup:

 

[Idontcare]

I'm a fan of Watson as well, its just not integrated as well as it could be for the benefits of the cmos development team, but if it was then it would be too focused as well, there are no easy trade-offs to be made...somebody has to do the CMOS pathfinding and the fab environment is not conducive to such laboratory-driven research.

I'm still divided on whether I wish TSMC to retain a vast majority of the pie, or if I want GF (or other foundries) to catch up and eat a lot more of the pie. On one hand, the foundry business is expensive, and having one clear leader means that company has enough funds (naturally, this means customers riding on them) to keep on improving process tech on a more timely manner. On the other hand, a more equal distribution between two players might not be so bad, allowing just enough funds to accommodate advancements in process tech without affording the luxury of slumping around. I lean more on the first, since as it is, foundries already can't afford to slump around if they ever dream of being (or remaining) relevant.

Everybody that is not a high-level TSMC employee or TSMC shareholder feels the same way. The entire industry, from the tool vendors to the consumable suppliers and even the TSMC R&D engineers depend on the cashflow that comes with corporate strategies that involve fierce competition.

As an engineer at TI the worst thing that happened to us was the day Moto spun-off Freescale and basically threw in the towel. TI went on to secure some 80% of the TAM for mobile phones and corporate decided it was time to cut costs everywhere they could in the name of shareholder and suddenly all the cool things about working at TI - the office parties, bonuses, conference attendance, etc - evaporated and dried up. Corporate bonuses got larger every year, corporate profits got bigger, TXN (stock ticker for TI) went up. But the general work environment started to suck and the money-train for the vendors began to dry up as well.

We could no longer get easy joint-development projects because of the required legal commitment of funds would not pass budget at corporate, vendors live and die by JDP's and JDA's.

So yeah, absolutely, with TSMC effectively squashing SMIC (didn't really move on past 130nm), IBM (foundry work died more or less at 90nm), Chartered (didn't really move on past 65nm), and now UMC (took forever to get 45nm out and they've gone on record as saying 28nm will be intentionally delayed for an even slower node cadence cycle)...GloFo was the industries hope for 28nm and beyond.

The challenge for everyone in the pure-play foundry business though is simply one of unavoidable accounting. If you sum up the total profits to date for the entire foundry business since its inception you come to the realization that TSMC has made 100% of all the profits.

Every other foundry that has been in business has done the AMD thing where some years they make money and then they lose it all over the course of the subsequent two years.

In that sort of an economic reality, the survivability and longevity of anyone and everyone that is not TSMC is continually scrutinized by the fabless customer base. And it becomes a self-fulfilling prophecy...because there is so much fear regarding the risk to the downside of choosing a lousy foundry partner, so many customers flock to TSMC that it deprives the other foundries of the very revenue stream they need in order to fund their R&D to avoid making it come true.

This is where the EETimes article comes in. TSMC would not have so many 28nm tapeouts if the reality on the ground was any different. If customers had confidence in GloFo's 28nm, or a lack of confidence in TSMC's 28nm, then we'd see GloFo publicizing this as well.

Unfortunately for me, as much as I've got a horse in this race for GloFo to come out ahead, I can't ignore the writing on the way that has been such a consistent indicator of foundry directions going into the past. There is nothing unique or special about GloFo that is somehow going to make their situation different from what happened to every other foundry in the business.

You can say they got deep-pockets and their backers would not let them fail but I'll remind whoever says as much that the same was said of SMIC (backed by the Chinese government) as well and that played out no differently.

Its not about selling every wafer you can make, its about making enough money to justify staying in the business of spending billions to develop next-gen process tech. Every other foundry has fallen off the node cadence that TSMC has maintained for one simple fact - their revenue failed to grow fast enough and large enough to justify the requisite R&D budget to stay competitive with TSMC. Sound familiar?

 

[Idontcare]

From that I now know the possible limit is around 5-10 nm and most probably move on to other materials for future processors.

Not a very well made pdf but the amount of information that it contains is awesome. Thanks :thumbsup:

If you like reading up on that kind of stuff you might find the links in this thread to be interesting/entertaining as well

 

[jimbo75]

The ramp rates on 40nm have hardly been an outlier compared to historic norms.

40nm has not been blockbuster, nor has it been dead on arrival.

They did have capacity ramp issues in the 3rd quarter as AMD made a lot of justified publicity regarding. But I haven't heard much being said about it since.

You can call it "the worlds worst process" but the argument here really is hollow as TSMC would not have been capacity constrained if no one wanted it. If it was really as bad as you want to claim it was/is then why are so many customers using it? And why are so many customers lining up to use their 28nm process?

TSMC's 40nm can't be all that bad if even AMD was willing to produce 40nm bobcat with it. Surely AMD wouldn't make such an egregious error as to go with "the worlds worst process" for debuting their first Fusion product, would they?

AMD had no option but go with it, same as the rest who are basically stuck with their business depending on TSMC not screwing it up again.

As for the graph, don't forget how they "solved" their yield issue which was basically throw more $billions at it by doubling capacity. If it had shown profit instead of revenue it would have been a long way down, in my opinion. Their 40nm has had unprecendented demand and should have easily been highest on that graph.

And there is yet another reason why customers are sticking with it. With bottomless pockets they know TSMC can do the same after they screw up 28nm as well. If this is what you meant with "confidence" then maybe you are right but I seriously doubt AMD or Nvidia for two will have much confidence in them after the yields debacle.

 

[itsmydamnation]

you also have to consider the backers of GF, they have money now but maybe not have as much in the long term future, they have already dropped billions of dollars into buying semi conductor companies so I doupt they will walk away. They have new york being built and i believe they have commited to building another Fab in Abu Dhabi, I dont think the financial success of 28nm Bulk is that important, but they need to start building a customer base somewhere.


edit: i also wonder if GF will put aditional funding into process development? Especally if they think Intel is considering selling its Fabbing services.

 

[Keysplayr]

Just read through this thread and I have to ask everyone. Do not take on thread disrupters on your own, even when defending a friend (which in and of itself is commendable). Don't do it here in the tech forums. Use the report post feature with a description of what you think is wrong.
I've just looked and see several posts that probably should have been reported but they were not. Be diligent in this respect. You want clean forums, use that feature. Not saying to abuse it either. Don't report somebody for merely disagreeing with you.

Now I'll ask all participating in this thread to stick to the subject matter ONLY. Do not discuss other posters or what you think of them for having an opinion. Also, if you have links to back up so called "facts" that would also be beneficial to the thread discussion.

Thank you.
Anandtech Moderator - Keysplayr

 

[Cerb]

Mmm, popcorn.

Alright, let me ask a question, regarding TSMC v. GloFo.

You are a higher-up in Anansi Products, LLC. You make controller SoCs for Creepy Storyteller Spiders(tm). Now, for whatever reason, you benefit from getting in on a cutting edge process early. Outside of high performance chips, I'm not sure what the reasons are, as I don't know too much about the business and tech involved. However, you benefit from being right on the cutting edge.

Now, let's say you're in this position, you have a history with TSMC, and you would not significantly benefit from the IBM/GloFo tech, over TSMC's. You don't use multiple fabs.

Why would you not stick with TSMC, unless they did something to screw up your business relationship with them?

Now, let's say you are in this position, but could take advantage of some of the IBM/GloFo tech over what you can get at TSMC. However, you aren't confident in GloFo. You also are not a huge company, so trying to use both GloFo and TSMC could put you in scary financial straits, if GloFo flopped, and you'd be even worse off if they flopped, yet you didn't have a backup plan. Sure, they'd get bought up, but in the meantime, you are out millions, and might not be selling enough good chips to make it worthwhile.

How would you decide, rationally, if it is worth it to go with GloFo?

Sure, there are scenarios that could more obviously steer someone their way, but what does the brand new untested fab have, basically offering the same thing, over the guys that have been around the block a few times?

 

[drizek]

There are 3 reasons to go with GloFlo

1. If they give you a lower price than TSMC
2. If they are ahead of TSMC
3. If TSMC doesn't have enough capacity.

GloFlo aren't going to leave their fabs sitting and collecting dust, and if #3 isn't enough motivation then they will have to go with #1 until they make #2 a reality.

Oh, and #4: You own half the company and have a vested interest in its success.

 

[Martimus]

There are 3 reasons to go with GloFlo

1. If they give you a lower price than TSMC
2. If they are ahead of TSMC
3. If TSMC doesn't have enough capacity.

GloFlo aren't going to leave their fabs sitting and collecting dust, and if #3 isn't enough motivation then they will have to go with #1 until they make #2 a reality.

Oh, and #4: You own half the company and have a vested interest in its success.

As for #4, I have some experience that would question whether that is AMDs motives.

When I worked at Delphi, we often made better quality parts and charged less than our competition for GM (for a great many reasons, not the least being that we had the facilities already set up and were once producing the parts as GM). This fact didn't matter, as GM was actively trying to break their dependance on us, and would buy more expensive parts from our competition, even though it didn't meet or exceed the specifications as well as ours did.

AMD could go the same route, and try to break their dependance on their old fabs. I am not saying this will happen, but it is a possibility. The dream scenario at GM was to build none of their own components, and have multiple manufacturers compete with each other to build the components for them. It worked for Chrysler to bring them back to profitability after they nearly folded, and GM and Ford wanted to follow suit to limit their own liability in the same way. AMD may be trying this same strategy.

Of course going this route has a great many problems as well. You don't have control over your own destiny this way, and there is no guarantee that you can find a supplier to actually build and supply the parts you need. In the foundry businiess this is even a bigger issue, as there are only a very small handful of foundries that can produce cutting edge high performance nodes.

EDIT: Also AMDs share in Global Foundries as now less than 34&#37; (although I don't have an exact percentage since the deal to purchase Chartered greatly decreased the AMD share in the foundry)

Source: http://www.xbitlabs.com/news/other/..._to_Acquire_AMD_Share_in_Globalfoundries.html

EDIT2: AMDs share in GF is now at 14%. Source: http://newstipster.com/atic’s-gain-in-global-foundries-is-amd’s-loss/224266/

 

[drizek]

I see two competing motivations for AMD, assuming nvidia will not, out of principle, work with GloFlo

One, AMD wants to keep manufacturing at TSMC to keep the pressure on nvidia. Especially if TSMC becomes scared of losing AMD as a customer, they might give them preferential treatment, giving Radeon an advantage in the market.

Or, they cut off TSMC, depriving them of significant quantities of revenue and hope that causes them to fall behind, leaving nvidia with inferior process tech.

Of course, I generally think very negatively(ie. "how can AMD best screw nvidia?" rather than "how can AMD make the most money?")

 

[jimbo75]

AMD/TSMC relationship must be getting extremely awkward, lets be honest. AMD sure owes nothing to TSMC, they've had the short end since forever compared to Nvidia's favouritism.

On the other hand, TSMC doesn't want Nvidia as their sole gpu customer. When did Nvidia ever do any of the ground work there? And they were falling over themselves to blame TSMC for Fermi being broken. It is in TSMC's best interest to keep them both competing.

IT relationships are getting to be extremely complicated. It's like a web these days whereas before it was simple 1 side vs the other.

Or, they cut off TSMC, depriving them of significant quantities of revenue and hope that causes them to fall behind, leaving nvidia with inferior process tech.

TSMC could sell AMD's allocation no problem. There would literally be a mile long queue for it so it's not that keeping AMD in favour with TSMC, it's almost certainly down to AMD being a company who knows how to do business in a proper manner - which is something that is still kept in very high regard in the East if not here in the West.

 

[drizek]

What about next gen consoles? I'm assuming they are going to come out at 28nm, right? These chips are going to be huge compared to current ones, basically an order of magnitude more transistors, right? That should keep TSMC and GloFlo busy.

 

[ViRGE]

What about next gen consoles? I'm assuming they are going to come out at 28nm, right? These chips are going to be huge compared to current ones, basically an order of magnitude more transistors, right? That should keep TSMC and GloFlo busy.

Would they use TSMC? I cannot recall any console chip other than NV2A being fabbed by TSMC. All the current chips are what, Samsung and IBM?

 

[Tuna-Fish]

What about next gen consoles? I'm assuming they are going to come out at 28nm, right? These chips are going to be huge compared to current ones, basically an order of magnitude more transistors, right? That should keep TSMC and GloFlo busy.

Why would they make a new generation? As consoles themselves make a loss, which the manufacturer tries to recoup by selling games, it's to the manufacturers advantage to extend the generation as long as possible. There won't be a new console generation until there can be a very clear and visible shift in visual quality between the new and old games. Given how pretty the games for the old consoles still are, we might simply not be there yet.

The only exception to this is Nintendo -- the Wiimania seems to have stopped, and they might build a this generation +&#189; console.

 

[RdBiker]

Straw man

That said, TSMC has made its move and it is now up to GloFo to make a counter move. You whipping up a bunch of nonsensical rhetoric (and yes, your posts in this thread have really been little other than nonsensical rhetoric as proven by the utter lack of facts, data, links, or anything else to establish your credibility on the matter) is not the same as GloFo refuting TSMC's claims.

I care about this because I've got a number of good friends, people I use to work for, who are intending on making a living in Malta and their livelihoods are depending on 28nm being something more for GloFo than what it is shaping up to be. And I know this because I've seen the same business cycles happen to SMIC, Chartered, IBM, and UMC.

If GloFo doesn't gain traction on leading edge CMOS then they are going to recede from the limelight as quickly as they entered it. That is the nature of this business and has been since 1987 when Morris founded TSMC.

It doesn't help that AMD is publicly non-commital on moving their TSMC business to GloFo.

Do you know (or have any guesses) how capacity constrained GloFo has been recently? I mean do you think it's possible that they have few customers because they simply don't have any free capacity? I read through the thread but didn't find a straight answer to this though that came up.
I remember reading that one reason AMD couldn't team up with Dell in the Athlon days because Dell would've needed all the capacity they had but I don't know if they had that much demand when these data were collected.

I was also wondering how small customers TSMC and GloFo can have? I know the big players order huge amounts of wafers per month but how small(ish) are the smallest quantities a company can order from a foundry?

 

[Idontcare]

Do you know (or have any guesses) how capacity constrained GloFo has been recently? I mean do you think it's possible that they have few customers because they simply don't have any free capacity? I read through the thread but didn't find a straight answer to this though that came up.
I remember reading that one reason AMD couldn't team up with Dell in the Athlon days because Dell would've needed all the capacity they had but I don't know if they had that much demand when these data were collected.

AMD was capacity constrained when all they had for Cu was Dresden Fab 30.

That is the time you are thinking of.

Then they built Fab 36 (also on Dresden) but by then they had lost so much momentum and market share that they essentially mothballed it and the fabspace sat empty for a long time. (quite a bit of it is unoccupied even now)

I don't recall anyone anywhere ever saying GloFo is/was capacity constrained.

Which is good, as a foundry you suck if you so poorly plan for capacity ramps that it becomes a problem that falls through to your customers. There is no industry that is immune to this.

TSMC sucked at it for 40nm which is partly the reason why Rick Tsai was replaced as CEO by Morris Chang (the prior CEO who retired to give up the position to Rick in the first place).

I was also wondering how small customers TSMC and GloFo can have? I know the big players order huge amounts of wafers per month but how small(ish) are the smallest quantities a company can order from a foundry?

You can order 1 wafer from TSMC if you like, it all comes down to price.

1 lot is 25 or fewer wafers. A lot of customers (20% by volume) are only ordering 25 or fewer wafers at a time.

Only a few customers, less than 3%, are ordering >1000 wfrs for a given product line. (Note this graph is not by customer but by product, AMD would represent many products here with all their different GPU dies and CPU bobcat products)

 

[jimbo75]

Confirmation of Abu Dhabi, not exactly the industries best kept secret but proof that GF are in this for the long haul.

http://www.eetasia.com/ART_88006430...9113,9518437967,2011-05-20,EEOL,ARTICLE_ALERT

 

[drizek]

Making chips uses an incredible amount of water, and the last two fabs to be announced are in Arizona and Abu Dhabi?

I guess since Abu Dhabi owns GF, they can be a bit biased, but why did Intel choose Arizona?

Edit: that was a rhetorical question. The answer is some combination of lax regulation, tax incentives, subsidized utilities, etc. all in the name of job creation. In the middle of the desert.

 

[jimbo75]

Plenty of sand too.

(also abu dhabi is surrounded by water, which is a surprise to most Americans )

 

[Lonbjerg]

Plenty of sand too.

(also abu dhabi is surrounded by water, which is a surprise to most Americans )

You forget that it's filled with sodium chloride...a major minus.

 

[jimbo75]

While that is a large drawback, the salt can be removed.

 

[Lonbjerg]

While that is a large drawback, the salt can be removed.

At a step price:
http://en.wikipedia.org/wiki/Desalination#Economics
Again...not a benefit.