[Beowulf] Vector coprocessors AND CILK
James.P.Lux at jpl.nasa.gov
Wed Mar 22 18:47:45 PST 2006
At 11:16 AM 3/22/2006, Vincent Diepeveen wrote:
>----- Original Message ----- From: "Jim Lux" <James.P.Lux at jpl.nasa.gov>
>>If it's a full custom chip, figure a "first chip" cost of $2M. (layout, a
>>couple spins, etc., but assuming you know basically what the chip is
>>supposed to do and how to do it)
>Most projects are indeed having a total cost of around $1M for such chips,
>including programming and design and package design for
>>I work with a fair number of very low volume but fairly complex chips
>>(intended for space applications, but not in Class S quality grade) and
>>they all seem to run about $5K to $10K each, which must be a sort of basic
>That must be something not sold in a shop then, but something intended to
>cream off the university world. Those universities waste money by the
>so for them $10k is affordable handsdown.
There are a lot of non-university, non-consumer small volume applications
out there. IBM PCs were in that category originally (at circa $5K
purchase price in 1981) and lots of businesses bought them for what, today,
would be trivial applications. My first computer (that I forked out the
cash for, as opposed to used) was a Z80 based machine with 48K that had a
FORTRAN compiler and it was around $5K, in 1978. Hardly a consumer item in
those days, but not a research curiosity either. In my case, the
investment was worth it because it allowed me to develop software for a
bigger machine (a CDC 7600) without burning up zillions of dollars in phone
calls and buying time on a timesharing system (at $25/hr, I seem to recall..)
Today, people needing things like test instruments would happily pay $5K
each for a chip that made some measurement possible. I recall seeing a
2048x2047 CCD sensor for astronomical use that was something like $80K,
each, in the mid 1980s, and that's hardly something that was going to sit
on the shelf collecting dust. In that case, the high price was because the
yield was so poor (huge die, and, as I recall, the challenge was in getting
the dark current low in ALL pixels). They probably had to make dozens to
get one good one.
Consider something like a high speed oscilloscope with a fast A/D in the
front end. There's some sort of custom ASIC in there that demultiplexes
the data into a bunch of parallel memory banks (since memory doesn't work
very well at 4 GHz). They might only sell a few hundred of the scopes, and
they cost upwards of $100K each, so dropping in a few multithousand dollar
chips isn't unreasonable, if only because it either makes something
possible at all (because fast needs physically small because of speed of
light issues) or makes it fit in a reasonable sized box.
When you get into things like aircraft engine controllers or radar
altimeters or ring laser gyros, you have another moderate volume
application (hundreds or thousands of units a year) where expensive chips
have a role: because their existence (at any price) makes the unit
possible. Likewise with certain weapons systems where you've got an
expensive widget (so the chip is a small part of the overall cost) sold in
small quantities with very unique design requirements.
>However for products and cards that you want to sell to ordinary people who
>simply want a bit better card, a price of $10k is too much.
Sure.. but my point is that there's a fairly substantial market out there
for chips that cost thousands of dollars, if they meet some peculiar need,
in a market that can support the cost: either because the chip enables the
product in the first place, or it replaces something more expensive that
It's not all government and universities providing jobs for worthy
post-docs as a form of "white collar welfare" (and, heck, they did that
before the IC was ever invented: I'm reading a biography of Robert Hooke,
and a lot of what went on in the 1600s sounds a lot like pork barrel
politics and finding jobs for favorite sons, in amongst which useful stuff
got done. Nothing ever changes<grin>)
In fact, the existence of the burgeoning consumer market has caused
problems for the low volume markets, because the chip manufacturers aren't
interested in supporting those markets, except in a "full custom million
dollar spin" sort of way. I can't count the number of times I've read some
article in a weekly trade magazine about some nifty new RF or digital chip
that might be useful, called up the mfr, and had, as the first question,
"So, how many thousand units a week will you be buying in the first 6
months?" What's an experimenter or prototyper to do? [ And this includes
before I started working at JPL, where the problem is even worse, because
we want space qualified low volume chips with documentation back to the
sand from which the silicon was made.]
FPGAs have helped a bit, but *big* FPGAs aren't cheap, and neither are the
design tools to use them. And, I'm not interested in trying to find old
74LS series TTL on eBay to breadboard with in my garage.
We're truly caught in a cleft stick here.. on the one hand, we can build a
supercomputer (by any casual observer standards) for a few thousand bucks
with parts bought entirely at WalMart, because of huge consumer volumes
bringing the price down. On the other hand, if our needs can't be met by
those high integration consumer products, the price to develop a solution
seems to have gone up.
Digression: Maybe it hasn't, though, back in 1978, I bought a $5000
computer with a hot rod CPU that cost $200.. today that would be a $600 CPU
in a $15000 dollar computer. Probably a bigger factor in the "feeling" that
things are more expensive for tinkering is that he opportunity cost of my
labor is dramatically higher today than it was 30 years ago, so.. $50 in
parts and 20 hours of labor over a weekend at $3/hr is more like $150 in
parts and $2000 in labor today. The factor of 10 increase in the labor
cost shifts the make/buy decision significantly.
>>price for them to build small runs where there's not a huge NRE. Things
>>like MOSIS (http://www.mosis.org/) (or Atmel's equivalent, the name of
>>which I forget) can be less expensive, but probably not for something of
>>this scale. $5K probably covers the cost of running the wafer, dicing,
>>testing, and putting it in a package, in quantities of <100.
>>So, to get the $50/chip cost, you need an order of 40,000-50,000 pieces.
>No no. 1000-5000.
$2M/$50 = 40,000 pieces, and that's just the NRE... actual manufacturing
cost, in that sort of volume is probably in the $10 range, so, of your $50
chip selling price, $40 goes to NRE and $10 to mfr cost, for 50,000 pieces.
You can run the numbers a lot of ways, but you've still got to divvy up
that million or two to get the first part out. Clearly, if you want a low
price per chip, then million unit sales are the way to go, so NRE
constitutes a smaller fraction of the total selling price (the graphics
>For 20000+ you can get the entire product including packaging down to way
Perhaps... you could get lucky and not need more fab spins... you could
have an existing product that only needs to be modified..you could have
brilliant chip designers who can design by hand and not need expensive
design tools.. It IS certainly possible to produce a custom chip for
less, but it's a pretty unusual circumstance, and probably not typical of a
modern company with offerings like ClearSpeed's.
More likely, a company starts with a smaller functionality, and builds the
market that way. For instance there are some smallish companies that make
custom motion control ICs that probably started with simple ASICs that
could be turned into first silicon cheaply (as in you could mortgage your
house to pay for it: aka "bet the farm") and they sold them at moderately
high prices ($50-100 each) so they could get positive cash flow with a few
That's a far cry from a worldbeater high performance computing chip with
millions of transistors.
>>But they're not going to even be able to cover a fraction of the
>>development cost for that. But, perhaps, if they are thinking about
>>"buying market share" with OPM (other people's money). It's been done,
>>more than once.
>If you want to earn back your development costs with 1 client,
>then you better stop producing such a product.
>Only money wasting governments want to pay that much.
You don't need to earn it back with 1 client (although there ARE
non-government clients willing to pay for such things)..
You can do it with hundreds or thousands, but it's certainly not a consumer
product at that point. And for non-consumer products, a $5K or $10K chip
isn't totally unreasonable, if it meets the customer's needs.
More information about the Beowulf