[conspire] Quiet, Freedom-compatible NAT/firewall/misc box?

[Rick Moen]

Quoting Don Marti (dmarti at zgp.org):

[my posting of a CompuLab Fitlet description and URL:]

> Wow -- that looks ideal, thank you.  Also has 2 HDMI displays so can
> set it up to also run Synergy as extra workspace, or something.  Guess
> I'll wait until they're back in stock, maybe set up a temporary NAT
> box with an old laptop until them.

It always struck me that old laptops ought to make fine ad-hoc appliance
hosts.  You get low power, quiet operation, and an onboard UPS.  Having
two NICs might take some work, depending.


Those CompuLab guys do brilliant engineering.  Maybe someone who's less
of a hardware amateur could explain to me why the Fitlet has an 8GB
ceiling on RAM.  Dana alluded upthread to some 'Avoton' Atom CPUs maxing
out at 64GB (vs. 8GB) because they were designed to be server-grade.[0]
(I'm not hopeless at understanding this stuff, but am catching up on new
x86_64 hardware after ceasing ot do so around 2006.)

/me Web-searches.  2013 architecture review of the Avoton with block
diagrams and such:
http://techreport.com/review/25311/inside-intel-atom-c2000-series-avoton-processors
Very much better than all the lackluster Bay Trail 'Atom' stuff out
there.  

    it's capable of true 64-bit addressing

Yeah, that.  A lot of online discussions of newish hardware burble at
length about how many 'lanes' of PCIe and how many channels are
available to address SDRAM, but I keep thinking 'Wait, wasn't one of the
key benefits of x86_64 over i386 the expansion of RAM address space from
4GB to some reasonable subset of the theoretical 16 exabyte linear
address space (2^64)?  

What I remember is that x86_64 real-world designs were supposed to have 
48-bit pinouts, and 2^48 is 256 terabytes.  So, machines with a 8GB
limit (like the Fitlet, or most Atom-based machines) means some
cheapskate has limited some key hardware component -- CPU address pins
and decoder circultry? -- to just a _single_ bit more than what an 
i386 box could do.

The 'Avoton' Atom CPUs would thus seem to be achieving 2^36 (64GB) RAM
addressing.


I'm jetlagged at the moment, so I'm likely missing something (and won't
swear to my math being exactly right), but, well, if what all this means
is the pernicious effect of price pressure, that's kind of sad.  RAM is
cheap, and buying the ability to expand it is thus cheap future-proofing
(not to mention virtualisation).

The Fitlet is so terrific that I hesitate to cavil over something like
its 8GB RAM limit, but, knowing that the AMD A10-6700T SoC and others of
its Beema/Mullins 'Puma' architecture class can address 32GB (I
_think_?), it seems a shame they didn't go higher.

I might be wrong:  8GB might be the max you can get with 'Puma'.  I do
know that that ASRock / Kabini ('Jaguar' architecture) bundle I cited
from Newegg[1] can handle 32GB -- and 'Puma' (Beema/Mullins) was the
2014 successor to 2013's 'Jaguar' (Kabini/Temash) architecture.



[0] Everybody loves ECC for 'server-grade' gear, except for when you
have to write the cheque.  Personally, in my long experience, I've found
that the Linux kernel tells you very, very clearly when you have bad
RAm, though patterns of segfaulting and spontaneous rebooting that are 
unmistakeable.  Given that, and choosing to not run a more oblivious OS
such as MS-Windows, why blow the money on ECC?  Not worth it, in my
view.

[1] http://www.newegg.com/Product/Product.aspx?Item=N82E16813157518 
Note $79 for motherboard and SoC.  If you suspect I'm leaning towards
buying one of these bundles plus a compact mini-iTX case, 32GB RAM, and
an SSD or two, you'd be correct.