home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!uknet!mucs!mshute
- From: mshute@cs.man.ac.uk (Malcolm Shute)
- Newsgroups: comp.arch
- Subject: Re: uniprocessor design ceiling
- Message-ID: <7125@m1.cs.man.ac.uk>
- Date: 6 Jan 93 11:59:03 GMT
- References: <2340@sousa.tay.dec.com> <1ho4g1INNgac@nigel.msen.com> <1993Jan3.215458.7960@athena.mit.edu>
- Sender: news@cs.man.ac.uk
- Organization: Dept Computer Science, University of Manchester, U.K.
- Lines: 45
-
- In article <1993Jan3.215458.7960@athena.mit.edu> solman@athena.mit.edu (Jason W Solinsky) writes:
- >If (as I have been led to believe) there
- >exists a linewidth which is large enough for yield to be effectively 100%,
- ^^^^^^^^^^^
- You quite rightly recognise that it never exactly reaches 100%...
- it's just an exponential curve which at some arbitrarily defined 'knee'
- you decide to treat it as being negligably close to 100%.
- So... if you're going to do something drastic, like multiplying the number
- of process steps by a couple of hundred, you will not be surprised to find
- the position of this knee moving.
- [I've never really understood how people manage to read 0.7V as the position
- of the knee of the exponential function which is supposed to characterise
- the IV curve for a silicon PN junction... you only need to get a magnifying
- glass out, and the point at which the curve really starts to break away from
- the I-axis moves :-]
-
- >why hasn't anybody attempted to build a very tall, large line width chip?
- >Is the problem a difficulty in creating monocrystaline Si above the substrate?
- >Are problems with planarization cumulative such that working with too many
- >levels becomes impossible?
- -k
- It is tempting for me to write: Y = ( 1 + A.D/k )
- where: Y = circuit yield
- A = circuit area
- D = fault density
- k = number of process steps
-
- It is one of those theoretically-reasoned models which abound in yield modelling.
- Unfortunately, Stapper indicates that, tempting though its arguments are, it
- bears no correlation with any experimental results that he has gathered.
- Instead, he, and most in the industry, appear to favour the negative binomial
- model, which appears: -a
- Y = ( 1 + A.D/a )
- Where 'a' is an empirically measured 'clustering coefficient'. It is going to
- have the effect of rising to kill your 1000-processing-step 3D process, but not
- in quite the obvious linear way that the Y=(1+AD/k)^(-k) would have suggested.
-
- [If you want to persue Stapper's excellent review paper on yield models, it is:
- Stapper, C.H. (1989).
- 'Fact and fiction in yield modeling',
- "Microelectronics Journal",
- Vol. 20(1), pp 129-151.]
- --
-
- Malcolm SHUTE. (The AM Mollusc: v_@_ ) Disclaimer: all
-