another question

[John R. Mashey]

On Oct 11,  8:49am, David Hough wrote:
> Something that I have been wondering about in relationship to the PowerPC
> discussion is:  what effect does volume in embedded systems have on pricing
> in PC's/workstations?    Obviously the effect is major if the exact
> same chips are
> used, but I had the impression that usually different chips are used -
> e.g. different clock rate, or
> leave out the floating point, or add specific logic for the embedded
> application, etc.Obviously i960 volumes have no effect on i486 pricing,
> regardless of whatever technology transfer may occur internally in Intel
> with respect to design and fabrication.    I don't know if these chips are
> built on the same fab lines - I presume multiple slightly different chips
> on the same fab line reduce the cost of all of them somewhat -
> I think I've heard this as an argument for SPARC single-vendor per design
> over MIPS multi-vendor business approach.
> So what is the correct argument that embedded systems volumes enable cheap
> PC's/workstations?
>
> If you care to expound on this, you might as well send it to
> scientific-interest and numeric-interestavalidgh.com.

OK, although I'm not a real expert, and this is somewhat simplified.
1) Intel is an odd case, in that they can sell more than they can make of
incredibly-priced chips [X86s].  To some extent, i960s are a diversion,
and sometimes have to fight for their lives for wafer starts.  [People
like to allocate wafer starts to things that get highest profit/wafer;
486DX2s make a lot more $/wafer than i960s.]  This, of course, is why
pentiums are in short supply: from typical yield calculations, one expects
that with 6" wafers, in .8micron, 3-metal processes, you get 20 486DX2s for
every Pentium, i.e., every Pentium forgoes ~$8K in profit!  (and if they
ship 100,000 of them, that's $800M in profit gone, which even Intel notices.)

2) For everybody else:
	a) For a fab, you need to spend $$ to develop the process, build the
	fab, and fill it with equipment.  You must amortize the cost of all
	that across whatever mix of chips that you ship over the lifetime of
	the fab.  These days, a new fab costs >$500M.
	b) For workstation/server chips, the life of  fab isn't very long, i.e.,
	people are always looking for the newest, zingiest process, so maybe
	you get to use that process for high-end chips for 18-24 months.
	c) If all you make are workstation chips in your fab, those chips have
	to bear the entire cost.  This is why DEC's Palmer says they have to
	make 4-5M Alphas/year to pay for all of this [because the DEC fabs
	don't do a lot of things that aren't VAXen and Alpha chips.]
	Under such circumstances, it is clear that quoted prices and actual
	costs may well be unrelated, and it is very hard to tell what's going
	on inside a captive fab.
	d) Some RISC chips are fabbed in standard fabs that run many other
	kinds of chips, like SRAMs, fast logic, et, all of which help pay for
	the fab.  In order of goodness, here are the kinds of parts to have:
		- Microprocessors for which there is big demand and that are
		   cheap to make [i.e., 486DX2s].  IT is especially good to
		   have a monopoly or nearly one.
		- High-end microprocessors, in general [good $/wafer]
		- Embedded micros, or other non-commodity parts
		- Commodity parts [like standard SRAMs, DRAMs, etc]
	The way you run that kind of fab is:
	a) Keep the fab full all of the time, at worst by filling it up with
	commodity parts, as long as the price you can get exceeds the
	incremental cost of running a wafer thru it.
	b) If somebody wants a bunch of non-commodity parts, that's good.
	c) Vendors like embedded-control micros as follows:
		1) Sometimes they are workstation parts of the previous,
		or previous-previous geenration, and hence don't need special
		designs.
		2) Sometimes they are derived from those parts, hence their
		design costs are minimal - look at IDT's sequnece, for example,
		of R3000, R3001 [lower cost version with a few tweaks],
		IDT R3051/3052 [add on-chip cache, other features], IDT R3081
		[add floating point to 3051], IDT R3041 [squeeze 3051 down
		lower].  Another one is to see an SGI INDY turn into an
		arcade game in 1994, and a home Nintendo in 1995, ending with
		low-cost (but very high-volume;  Nintendo has 100M installed
		hardware systems (!)) variants of our low-cost desktop chips..
		which will also reduce costs for desktops (as long as they're
		not games!)
		3) A lot of software already exists for compilers, libraries,
		etc.
		4) REALLY IMPORTANT: these things tend to sellable in useful
		volume for years and years, and *dno not* need the latest
		process, and yet get more $/wafer than commodity stuff,
		although not the premiums that the newest chips command.
		Nevertheless, they extend the economic life of the fab ...
		which means that the cost is spread over a longer time.
		5) ALSO IMPORTANT: if you are a chip company, and you have
		CPUs, you get to talk to customers, and also maybe sell them
		other support chips, and you can amortize the cost of sales
		and support across a variety of chips.  If you don't have a
		CPU, sometimes you don't get to talk.  If you *only* have CPUs,
		(with exception of X86-land), it's extremely difficult to
		sell and support large numbers of design-ins cost-effectively.
		5) However, to be successful, an embedded control line needs
		a large number of variants, and must cover a useful cost
		range [there are MIPS chips as low as $15 already, and
		embedded version go up to about $200, I think, seldom more,
		except occasional high-end applications that use workstation
		chips anyway].  Anyone who thinks a $300 chip is "low-cost"
		in this market hasn't been in it!

3) A few miscellaneous notes by RISC vendor, in order from:
	USES MULTIPLE STANDARD CHIP VENDORS ... TO ... CAPTIVE FAB ONLY
	(SOME OF THIS IS POSSIBLY-BAISED OPINION).

	a) MIPS does base designs, usually for high-end micros, but sometimes
	for other ones [i.e., R4200 is low-power/cost oriented].  We use
	vanilla industry fabs, with no unsuual tweaking [which chips vendors
	generally won't do], and we always make sure a new high-end design is
	buildable on at least 2-3 of our partners' fabs, so there are multiple
	sources for the base chips.  Partners often do derivative designs:
	there are at least 15 variants based on R3000s, including ASIC
	versions (like LSI Logic's, and I guess, now public, what Sony is
	doing for consumer products, i.e., a custom/synthesized core plus
	a chunk of gate array for customization].  Making these variants
	possible requires  specific work in tools, data formats, etc akin to
	making UNIX more portable in the first place.
	MIPS chips are actually quite noticably used in embedded control
	[there's supposed to be about a thousand design wins, counted the way
	chip people count; in image boards, airplanes, telephone switches,
	printers, etc, etc.]  I'd guess, in 1993, there should be >1M
	MIPS chips, of which maybe ~10-15% will be in computer systems (UNIX,
	Tandem Guardian, Win/NT) and the rest in embedded control of one
	sort or another.  Sooner or later, of course, consumer products
	dominate the unit volumes [TV set-tops, Nintendo games, Sony
	organizers/videogames/later PDAs (maybe).]
	A bizzare rumor is that MIPS' use in embedded control was some kind
	of accident.  This is wrong.  The explicit requirements were that the
	architecture be good for UNIX, but not UNIX-specific, and be OK for
	high-end embedded control (at first), and lower (as costs came down).
	Having helped design the relevant features [MMU, exception-handling],
	I *know* that embedded was one of the goals, and I also spent a lot
	of time with some of our semiconductor partners
	to help that area get going.
	(It's now mostly handled by our semi partners directly.)
	Note, of course, that one must consider both $ revenue and unit
	volume to look at investment needed.
	Note that the base chips will always have multiple sources;
	derivatives may or may not [for example, some of IDT's designs are
	second-sourced thru Toshiba and Siemens.]
	Finally, fabs cost multiples of $100Ms, designs cost multiples of
	$10Ms... but hte semiconductor partenrs are going to invest the $100Ms
	in any case, as MIPS chips are a modest fraction of what they do.


	b) SPARC: as far as I can tell, most (but not necesarily all)
	successful SPARC designs have been done by Sun, generally working with
	a single chip partner [i.e., with Fujitsu on the gate arrays, with
	Cypress for 601, with TI on SuperSPARC, etc.]  and the focus has
	usually been on designs for Sun's own needs [not surprisingly].
	As far as I can tell, there aren't a lot of SPARCs in embedded
	designs (except maybe for use in VME boards and such, what I mean is
	not a lot of design-from-scratch applications.] I'd be interested to
	hear of any such public ones to add to my lists.
	AS far as I can tell, most currently-used SPARC chips do not have
	pin-compatible second-sources.  [Older ones sometimes did.]

	c) HP: uses internal fabs, but has done foundary work for others
	[to fill up the fabs, of course, as described above], but is getting
	out of that business as theire capacity needs are now using up their
	capacity.  New fabs cost enough to develop that they're sharing the
	process development effort with AMD for the .35micron round, I believe.
	AS far as I can tell, HP PA so far is mostly used in HP systems,
	with a few others (Samsung, Hitachi) doing some. I don't know of
	any serious presence of HP PA in the embedded market yet.  Hitachi
	designed an embedded version a ew years ago, but it didn't seem
	very competitive.

	d) IBM / Motorola.
	RS/6000s are only made in IBM-internal fabs.
	PPC 601s are only made in IBM-internal fabs [use 4-metal process].
	Later PPCs will be makeable in both IBM and Moto fabs.
	IBM has reputation for absolutely superb fab technology, but expensive
	(hard to tell, since can't see internal accounting;  price and cost are
	sometimes difficult to relate when looking at chips made only in
	captive fab).
	IBM, of course, makes many kinds of chips in its fabs ... but really
	getting sustainable prices down will probably be for Motorola.
	So far, RS/6000s are mostly limited to systems [IBM Federal actually
	uses MIPS R4000s for embedded avionics, for example].  PPCs are more
	reasonable for embedded control, although a lot of embedded could
	care less about floating point.
	Moto will do embedded versions, of course [and I've seen press on IBM
	semiconductor group doing something called the 400-series, which it
	appears to be promoting (in competition with Moto??).]
	I think all this means there are 2 sources for the mainline PPC family
	after the 601, and otehrwise can't tell.
	e) DEC Alpha -  (as mentionend above)
	uses captive fabs, that mostly make VAXen and Alphas.
	Palmer says they have to ship 4-5M/year to pay for this, hence the
	reason for wanting to do lower-cost, higher-volume chips for mobile,
	embedded, etc.  [From experience, this is going to take a long time;
	a lot of the market wants low-cost ($20-$50) or low-power, or both.]
	They use CMOS with a bunch of tuning, which made it difficult to
	get a second source [they asked *everybody*, finally, Mitsubishi said
	Yes, for late next year].  Chip vendors want to see volume, and do
	*not* want to run their fabs with knobs and dials turned to one side,
	as it usually hurts yield.  [Note: if I were DEC, I'd probably have
	done the same thing, i.e., since have under-utilized captive fab, go
	all-out for performance with heroic circuit design, as lower yields
	are irrelevant at this time.]
	I'm not sure I understand the long-term economics of all this; fabs do
	cost real money.

4) A few numbers:
According to RISC Management (but cross-checked with otehrs, and probably
close), in 1992:

Intel i960	2095K units	(HP LaserJet; more in 1993)
AMD 29K		 849K		(laser printers, more in 1993)
SPARC		 320K		(more in 1993, mostly from Sun growth, I think,
				as Sun was something like 90% of SPARC w/s)
MIPS		 291K		(about 90K systems, 200K embedded)
				Expect >1M in 1993, from OKI printers
(these are merchant chips above).
HP PA looked like in 70-90K range, and IBM in 60K range, in 1992,
although I didn't have one singel source for those numbers.)