[IAEP] Moore and the OLPC (was: The Guardian: PlayPower: 1980s computing for the 21st century)

[Jecel Assumpcao Jr]

Martin Langhoff wrote:
> > The main factor for the low costs is Moore's law: you
> > can either get twice the transistors for the same price in 18 months or
> > the same transistors for about half the cost.
> 
> I don't think it's quite like that. Making chips is only cheap if you
> have huge volume. Basic QA of chips and boards is costly. Assembly (it
> has countless parts) is costly too. Financing a production run
> requires a lot of money, and stocking all of that costs... lots.
> 
> It's very likely the they are just selling very old stock -- that's
> the only thing that'd explain the price. As soon as it runs out, any
> crazy entrepreneur that wants to make more will find out the real
> costs.

and Yama Ploskonka wrote:
> Elonex One clones are available right now for about $75 USD in
> quantities over 100.  They were released original well after the XO 1,
> and have about similar hardware.  Originally they sold for about $300.
>
> The XO seems to be about the only one defying Moore's :-)
> 
> While the (heavily subsidized) pricepoint of $200 was totally amazing
> in 2007, right now it is rather unimpressive

I oversimplified the issue the last time, so I will go into more detail:

Moore's Law applies to equivalent circuits built using different
generations of chip manufacturing technology.

The price of a single circuit also varies with time due to what we might
call "the learning curve". When you first introduce a chip it is in the
"state of the art" phase and has a high initial cost. The cost of
development (known as "non recurring engineering costs or NRE) must be
recovered from the initial sales and the chip is probably the first of
its kind so the customers don't have much choice but to pay to get this
functionality. In addition there are typically some initial problems in
the factory and the working chips have also to cover the cost of making
the bad ones.

As production ramps up you get some economies of scale, higher yields
due to experience, finish paying off the NRE cost and start to feel the
pressure from competitors so that the price of the chip comes down in a
big way. This is totally unrelated to Moore's Law since you are using
the same factory and same chip technology. We can call this the "mature
phase".

Prices don't come down forever, unfortunately. As the market's focus
moves on to newer chips, the old one tends to become more expensive as
it is a bother to the factory to keep making it for the few customers
who still need it. Often the product is simply pulled out of the market
entirely. This is the "obsolete phase".

So the aspect of the OLPC strategy that took advantage of Moore's law
was building a 1997 spec laptop in 2007 using newer generation chip
technology to make the old design (the Geode was adapted to relatively
new AMD processes). The aspect of the OLPC strategy that was supposed to
take advantage of the learning curve was to use mature products instead
of state of the art ones and to have such absurdly huge volumes ("we are
thinking of making a few billion....") that it would force component
prices to keep going down rather than being eliminated or becoming more
expensive. Of course, if you design your own chips then they are
necessarily "state of the art" so this was against the initial plan, but
in the end reality forced the two custom chips into the machine.

Since the volumes were far lower than planned, this second aspect
failed. AMD stopped making the GX Geode, then all Geodes. DDR memory
became more expensive than newer ones. It is practically impossible to
get the Flash chips that were originally used. You can bet that the
companies involved would have acted differently if given orders of a
hundred million units. So the idea was good, but depended on the
customers not taking years and years to make decisions.

The current OLPC strategy is to take advantage of the constant
price/increasing features curve of Moore's Law. So you go from the 128MB
Geode GX in the XO-1 B1 to 256MB Geode LX in the final XO-1 to 1Gb Via
in the XO-1.5 to the announced ARM in the XO-1.75. You get better and
better machines for a little less than $200. The two drawbacks of this
path are that there is the temptation of focus software development on
the latest model (leaving existing deployments stuck in old versions)
and the constant price limits opportunities for new deployments.

To address the comments: it is very likely that these cheap Famicom
machines are not old stock but are in production right now (the Asian
market is pretty crazy). The Elonex One is taking advantage of the
learning curve with a slightly better choice of components. Actually,
the learning curve for the OLPC was masked due to the large initial
donations. A normal startup would have had to sell the first low volume
batches for about $500 and only later would the price come down to $200.
Instead of that, OLPC even gave away the first few thousand machines for
free.

-- Jecel