Globalfoundries on Track for 50% Natural Yield by Year End with 32nm Process Technology

[Idontcare]

Globalfoundries on Track for 50% Natural Yield by Year End with 32nm Process Technology

Globalfoundries expects to start volume production of 32nm-SHP (Super High Performance) technology at Fab 1 in the second half of 2010. This technology will employ silicon-on-insulator (SOI) substrates and utilize Globalfoundries? innovative ?Gate First? approach to high-K metal gate (HKMG), which maximizes power efficiency and transistor scaling while minimizing die size and design complexity when compared to the alternative ?Gate Last? approach. Yield progress continues with 24Mb SRAMs in double-digit natural yields on path to 50% natural yields by year-end.

?When compared against the 45nm-SHP technology we?re currently running in Fab 1, we?re seeing performance improvements of up to 50% in the 32nm generation at the same leakage levels of the 45nm generation. When you combine this with our patented Automated Precision Manufacturing (APM) technology and exceptionally low defect densities, we believe we?ll be in the leading position among foundries to bring this technology to market in volume for our customers,? said Jim Doran, senior vice president and general manager of Fab 1 at Globalfoundries.

http://www.xbitlabs.com/news/o...rocess_Technology.html

edit: fixed broken link

 

[drizek]

(I am pretty ignorant about this stuff)

24Mb SRAMs are tiny, right? That means yields for CPUs would be far worse than 50%.

Another question, lets say your 45nm yields are 80%, and your 32nm yields are 50%. Does that mean it would actually be cheaper to shrink to 32nm at this point(roughly speaking, of course)?

 

[ilkhan]

Might want to adjust "year end" to "end of 2010".

 

[Ben90]

Fuck 32nm, im done with computers! Gonna go get me a power kite!

 

[ShawnD1]

Originally posted by: Ben90
Fuck 32nm, im done with computers! Gonna go get me a power kite!

I tried to report that post to a moderator but I don't even see a button to hit for that. Am I missing something?

 

[Idontcare]

Originally posted by: drizek
(I am pretty ignorant about this stuff)

We all were at some point in our lives, until we asked questions, so never feel like you have to excuse yourself for it. At some point in my life I knew even less about this stuff than you do now, only one way I got from there to here - asking lots of questions.

Originally posted by: drizek
24Mb SRAMs are tiny, right? That means yields for CPUs would be far worse than 50%.

yeah reporting yields on a 24Mb sram chip is not exactly confidence building...that is only 3MB after all.

Consider that 45nm PhII has 8MB of sram. So if you are getting only 50% yield on just 3MB of sram you can figure the yields on a CPU is going to be pretty much zero at this time.

(no surprise there on the zero yield for cpu thing, this is typical of all process nodes as they are iteratively improved upon during the development cycle...what is "new" here is that GF actually publicly divulged this info - a good thing - but in doing so they stated the test chip qualifier - 24Mb sram - and that's kinda embarrassing of them to publicly acknowledge they use such a teeny-tiny easy to yield test chip)

Let's see...24Mb at 0.149um^2/bit means the cells themselves occupy around 3.6mm^2, add say 10% for the write buffers and round it up and lets just call it 4mm^2.

Yielding 50% on a 4mm^2 test chip is not the kind of info you want to give your competition's marketing guys. TSMC and UMC are going to make hay with this info.

I feel for them, I know what it takes to get corporate to allow you to divulge yield info so I'm sure folks worked very hard to make this info public with hopes of it boosting confidence in GF's process prowess, but they really should have withheld the technical details of the test chip size as that info actually undermines their efforts (IMO).

Maybe I am misreading something here? Is it a typo, are they talking about a 240Mb test chip instead?

Originally posted by: drizek
Another question, lets say your 45nm yields are 80%, and your 32nm yields are 50%. Does that mean it would actually be cheaper to shrink to 32nm at this point(roughly speaking, of course)?

In a manner of speaking, yes that is how these "cross-overs" are handled. Each node raises the cost per wafer, in general, but about 15-20% versus the cost of producing a wafer on the prior technology node.

The cost increase is generally acceptable because the new node enables you to make more chips per wafer, or can be viewed as enabling higher ASP products such that your gross margins cover the elevated cost structure.

So yeah, somewhere in there if you add enough caveats and state all the fine print about things being considered equal and other things held constant in regards to what it is that is yielding 80% at 45nm versus what exactly is yielding 50% on 32nm, there is a yield-cost equivalency point at which it is "six of one, half-dozen of the other" in terms of the costs to make a chip on one node versus the other...however cost is only one factor in determining gross margins (the thing they want to maximize), the other side of the equation is ASP.

Long before we get to cost-equivalency we probably reach a point where the newer node enables products to perform at a level that generates an ASP that is higher than that of the ASP enabled by the performance of an older node's process technology.

That 32nm chip which can be sold for $200 but costs $50 to make versus that 45nm chip which can't be sold for more than $150 but only costs $30 to make...which one has better margins? (the 32nm one in this example of course)

Yields (cost) are just one part of the equation, ASP entitlement is the other relevant part of the equation.