home *** CD-ROM | disk | FTP | other *** search
- From: sjmz@otter.hpl.hp.com (Stefek Zaba)
- Date: Mon, 23 Nov 1992 21:27:28 GMT
- Subject: Re: Apple Crippling Systems?!?
- Message-ID: <72140006@otter.hpl.hp.com>
- Organization: Hewlett-Packard Laboratories, Bristol, UK.
- Path: sparky!uunet!elroy.jpl.nasa.gov!sdd.hp.com!hpscit.sc.hp.com!scd.hp.com!hpscdm!hplextra!otter.hpl.hp.com!otter!sjmz
- Newsgroups: comp.sys.mac.misc
- References: <1992Nov21.020604.224@physc1.byu.edu>
- Lines: 38
-
- In comp.sys.mac.misc, avery@ccrma.stanford.edu (Avery Wang) writes:
-
- > .... It's hard to imagine a chip design failing for the FP operations being
- > not "up to spec". After all, it's not an analog circuit. And digital logic
- > is, by definition, up to spec. So it would have to be intentional, I surmise.
-
- Oh, if only! If only! Down at the electrons on the chip doing their pretty
- quantum dance, it *ain't* usefully "digital". It's noisy, ugly, stochastic,
- statistic. It isn't (with apologies to Stanislaw Lem) a dream of a machine
- at all. There's a range of voltage levels we choose to call "high", a range
- of levels we choose to call "low", and a spread in between during which the
- signal's rising and falling which all good chip designers PROMISE ON THEIR
- GRANDMOTHERS GRAVES not to sample. Chip layout is more and more of a black
- art the closer you are to the bleeding edge, and Motorola got as least as
- close to that edge as they wanted to with the '040. Add to that the unwanted
- impurities and crystal defects you get in silicon wafers, and you understand
- why chip yields are *NOT* exactly 100%, and why the same semiconductor
- fabrication line can produce an 80% yield of 80ns memory chips one day, and
- next day be down to 60% at 100ns only. (I exaggerate for effect; nor are
- these yields suggestive of typical experience. First run chips have been
- known to yield 5% or lower, and it certainly gets worse as the chip area
- increases, since you're that much more likely to hit a flaw in the silicon.)
-
- Hence it's feasible --- nay, common --- to test chips as they are minted,
- and put them in suitably labelled packages according to the test results.
- It also makes simple sense that a smaller surface-area '040 WITHOUT the
- large chunk of silicon devoted to floating-point ops will give a higher
- yield than the larger-surface-area full function '040.
-
- Final piece of well-known context: each wafer of silicon which goes through
- the fab line (say, 4 inch diameter) has multiple copies of the "chip" etched
- onto it --- it's then cut up into the individual chips. So, making "larger"
- chips doesn't mean using bigger wafers, but means putting fewer copies of a
- larger design onto the same wafer.
-
- End of "chip fab 101"... from a software weenie.
-
- Cheers, Stefek
-