KISS (fang.qiu@gmail.com) on 9/17/09 wrote:
---------------------------
>I mean what if developing an ultra-low-power server chip using ARM core(i.e., work
>done per dollar and work done per joule). Just like styles of these machines:
>BlueGene (powerpc core)
>SiCortex (mips core)
>SGI's Molecule (atom) (http://www.theregister.co.uk/2008/11/20/sgi_molecule_concept/)
>CMU's FAWN machine(http://www.cs.cmu.edu/~fawnproj/)
>And Microsoft research's Atom machine(http://www.theregister.co.uk/2009/02/25/microsoft_sleepy_servers/)
>The core idea is rebalance cpu core performance, power, memory system, i/o, interconnection,
>for new workload environment(such as web service...).
>
>what's the obstacle to develop such machine around ARM:
>1, lack of 64-bit is a problem?
>2, at software-side - does ARM ISA/Toolchain have good support for LAMP staffs?
>
>And will Google(whatever) license "embedded cores" and design a custom SoC, to
>reduce power/cooling cost in data center and save their electricity bill? Can the
>return can support the investment of a new silicon design (45nm or others)?

I've got no inside knowledge about Google, but I _think_ they tend to put quite a bit of CPU power in each system. High end (eg. 4 core) single socket, or even dual socket per system.

One reason would be perhaps power, footprint, and cost (for the other components) per unit of CPU power.

Another reason I think would be quantization of system resources. Given an amount of RAM, CPU, IO capacity etc, they try to fit complementary workloads together to maximise utilization of all these resources. Having less resources per system and more systems probably makes this less efficient (get more "fragmentation").

Power efficiency I'm sure is a very large concern, but absolute performance I think would remain quite important.