On 30/12/13 14:55, James Salsman wrote:
If you don't like Cubietrucks, then how about
RADXA? At least with
http://dl.radxa.com/rock/docs/hw/RADXA_ROCK_schematic_20130903.pdf
you know exactly what you're getting and it doesn't cost a huge power
bill.
Maximum 100 Mbps ethernet connection. Also, it doesn't exist yet.
We still failover when machines go out of service, and
sure the
caches would have different RAM configurations, but the fact is it
doesn't cost more money to switch to ARM, and you jettison a bunch of
legacy x86 crap that nobody uses but take millions of transistors
which need to be powered. Why ask our donors to keep all those useless
transistors warm?
Are there some benchmarks which support this idea? I read
<http://armservers.com/2012/06/18/apache-benchmarks-for-calxedas-5-watt-web-server/>
But it was full of distortions, like comparing the actual power usage
of the ARM system with the TDP of the Intel system, and then using a
workload which saturated the network link of the Intel system versus
the CPU of the ARM system. Maybe this sort of fluff is part of the
reason why Calxeda went bust.
Marketing materials on Calxeda's website indicated that the system was
priced such that it would be more expensive than Intel on a per-MIPS
basis, but that you'd win in the long run through reduced power bills.
It didn't sound like a cheap solution to me.
I read this:
<http://www.theregister.co.uk/2013/12/13/facebook_arm_chips/>
But it was clear that it was only at a prototype stage -- the
benchmarks are not in yet because the development work needs to be
done first. I read this:
<http://www.theregister.co.uk/2013/12/16/google_intel_arm_analysis/>
which speculated that Xeon may still be better for CPU-intensive
tasks, and ARM chips may be useful for storage control. But a
Cubieboard or Radxa can't be used for storage, since they lack the
necessary high-bandwidth connections.
Leslie Carr wrote:
At that point we'll probably need to redesign
those boards
which are incapable of doing these things, so we'll need a team of
hardware engineers, plus a deal with a manufacturing plant.
Google and Facebook are apparently taking that route. Maybe some day,
this technology will be available for anyone to buy.
-- Tim Starling