"People whine about foreign oil but foreign monopoly on chip production is just as big a threat."
You make an excellent point; I believe INTC's (surprise) increase in capex is probably geared towards 450 mm wafer processing and quoting the "Godfather" Intel might make Apple an offer they can not resist. 2011 was very unique year exposing risks related to offshore outsourcing.
I can think of Apple and Intel working together to create a best of breed SoC (manufactured at 450 mm). IMHO too many US based fabless chipmakers are competing for advanced technology capacity. An usually a foundry process is more or less a one size fits all process - it's probably good but not as perfect as it could be finetuning all parameters for one specific customer.
Yield data is usually top secret info but TSMC is ramping up BIG time (for TSMC standard) in
Q1 2012 - however TSMC added almost no additional 300 mm capacity in second half of 2011 according to TSMC quarterly report:
According to latest quarterly report TSMC added 92 units of 300 mm wafer starts in 2011; in Q1 2012 alone TSMC is adding 71 units of 300 mm wafer starts (Fab 12, Fab 14).
So TSMC is playing catch up but according to ASML who keeps excellent track of what is going on at customer base TSMC's capacity additions for 28nm are far away from what would be required to make 28nm main stream.
So I would not write off TSMC prematurely but I just can't see someone like Apple continue to depend on TSMC when 450 mm becomes mainstream.
And Samsung is becoming more and more a competitor for Apple.
450mm is a cost reduction measure that squeezes more die on a wafer. It has nothing to do with "technology" leadership. It's the transistor dimensions and die size that are important. 22nm & 14nm will be on 300mm before they bring up 450mm.
Apple picked up a "ton" of WiFi related IP from Freescale over the last six months or so - obviously it's patented but would you be comfortable to disclose this kind of knowledge with someone like Samsung who is direct competitor?
I certainly would not.
Jun 6, 2011 – Intel CFO Stacy Smith is on record saying that combining other chip design IPs with Intel's architecture core "would be fantastic business for us. ...
Well, if TSMC can't make a lot of good 28nm dice, then that probably explains why AMD's cancelled most of their 28nm Fusion products (Deccan/Wichita, etc.). Better to use them to make high ASP GPUs than low-ASP APUs.
"While problematic, TSMC have enough of a capital cushion to withstand such losses."
That is the problem. The big three have enough capital cushion for anything. They have no need for innovation except for their own. You have to turn to FTC to get anything, like in the Univ. of New Mexico vs. TSMC case. It is killing the ecosystem.
No mention of UMC 28nm in this article, but EE Times article from last year indicated TI using UMC as their lead foundry for 28nm. I assume other chip companies have similar multi-source foundry strategies.
It will be interesting to see which foundry can work out thier 28nm yield issues first and how this translates to application processor and GPU market share in 2013.
I think we may potentially see many fabless semiconductor companies talk about moving to UMC and/or Samsung for 22nm if TSMC has a setback at 28nm. But much of this will be idle talk to try and get TSMC to lower prices since nobody else has enough production capacity at this point. Most 28nm designs simply are not ready for production level tapeout since the process and IP are immature for 2011 and 2012. I suspect much blame also resides within TSMC's customers for the failure since just about everyone does test chips on new processes for characterization before launching production designs.
It looks like xilinx was really happy with tsmc 28 nm, so it's probably rushed designs that's the problem.
Since most of the large fabless companies are satisfied with TSMC 28nm process, the low yield issues of some companies are potentially due to design related process issues. It can be fixed by design/process expert after review the layout.
Samsung stays ahead of TSMC in this regard because they always start a new node with memory before going to processors. For memory Samsung is already at 20 nm. Samsung's A6 for Apple is already at 28 nm ( only in Austin, NOT So. Korea )
In the last few years TSMC has come through as being a little too cocky about their market share and at Technical Conferences sounded almost "Big Brother" ish.
Samsung's spending of 10s of billion $ in 2012 does not bode well for TSMC lead in Foundry. In fact TSMC is already out in the public back pedalling on their CapEx siting low growth in demand.
What happens now if nVidia or even Qualcomm ( current large customers of TSMC ) now jump ship ( Samsung, even Intel ?? ) in order to stay competitive with A6 Quad Core ?
It sounds like there is no clear story here, one says no yield and another says on target for defect density. It could mean the 0% yield is on track with what they expect for defects at this point. I am sure that they are closing guarding the real numbers. Only time will tell, I am rooting for success as it can only help drive costs down overall in the market.
Regardless of the truth of the details, 28nm has been a real challenge, particularly in the initial ramp up phase. This is more sanely evidenced in the increasing amount of DFM activities being pushed to the designer at these nodes. Things like litho checking, smart fill, pattern based checks, and restricted design rule checks have migrated from recommended to required. This is evidence enough that many of the second and third order effects on yield are becoming first order. Strict adherence and effort spent optimizing these DFM issues in the design phase is probably what is causing such different results on different chips. All designs are not the same from a DFM perspective.
My Mom the Radio Star Max MaxfieldPost a comment I've said it before and I'll say it again -- it's a funny old world when you come to think about it. Last Friday lunchtime, for example, I received an email from Tim Levell, the editor for ...
A Book For All Reasons Bernard Cole1 Comment Robert Oshana's recent book "Software Engineering for Embedded Systems (Newnes/Elsevier)," written and edited with Mark Kraeling, is a 'book for all reasons.' At almost 1,200 pages, it ...