[Arm-netbook] The future of EOMA-68

[Luke Kenneth Casson Leighton]

On Sat, May 2, 2015 at 10:44 AM, Paul Boddie <paul at boddie.org.uk> wrote:

>> What is the future of EOMA-68? Any EOMA-68 with a powerful hardware (like
>> Tegra X1 or Intel Bay-Trail)?
>
> Another consideration is openness. Are either of these technologies
> sufficiently open? Nvidia have traditionally had a bad reputation for this,
> perhaps only courting openness when they've struggled to attract customers, as
> I remember being the case with their SoCs: I think the summary was that they
> promised a lot and delivered comparatively little, and the customers all
> switched their future designs to other SoCs in disgust.

 so, in essence we have:

* NDAs which prevent and prohibit access to critical information,
  making even assessment of the SoC much harder than the
  competition

* Non-free binaries and firmware which, from bitter experience,
  are known to cause such complete and utter aggravation,
  support issues, upgrade and reliability issues that it's likely to
  cause such harm to a customer's reputation that end-users will
  reject the entire customer's products outright (look at what
  happened to OCZ as an example)

* pricing that's far too high for what it is esp. compared to entry-level
  successful readily-available mass-volume china-based SoCs

* features that are so far advanced of "good enough computing"
   that nobody notices the difference *anyway*

* not that many customers anyway because there's not that
  much buy-in (chicken-and-egg problem), so the MOQ has to
  set at 1,000,000 units... thus *increasing* the chicken and
  egg problem rather than decreasing it...

all in all, miguel, it's very hard to justify the expenditure of time
and effort, and there are in fact sound business reasons why the "more
powerful" SoCs should *not* be used - it's simply too much of a risk.

 at some point, someone in a non-china-based company *is* going to
"Get It".  they'll produce a long-term (5+ year) SoC, with between 300
and 600 pins, at a price between $3 and $5, that is at least quad
core, and, because it will be in 14nm or 10nm, will be absolutely
great performance and extremely good value for money.

 the problem that any such company faces however is that the yields on
anything below 28nm are beginning to get very dicey.  28nm can easily
manage a 95% yield, and at this low end of the market you can easily
get 4,000 ICs on a 10in wafer.  one wafer costs $USD 4,000, yield is
95%, packaging is $1, testing is $1, now you start to see why
Allwinner can sell a quad core 64-bit processor for $5: they're making
$1 profit per chip!

 by contrast: intel puts such vast amounts of cache onto their
processors (2mbytes instead of 128k for an ARM SoC) that they can only
fit say 500 ICs onto a 10in wafer... one wafer still costs $USD
4,000... you're now looking at a $12 per IC.

 *but*, let's say that they use 22nm, now you have 50% more ICs...
*but*, the yield is say only 60% because it's a lower geometry.... the
cost now just went *up*, not down...

 at 14nm the yield is only say 40% but you don't actually get 2x the
number of ICs because you have to allow for space between them in
order to cut the individual ICs off the wafer.... cost went up even
more...

 ... start to make sense?

l.