Yet another shout-out of respect to GigaOm, and particularly Stacey Higginbotham, who has posted twice about the potential move to ARM chip architecture in servers.
The first post is about Smooth-Stone, a start-up that is using ARM chips (widely used in phones and other mobile devices) instead of x86 silicon in a server, not for the high performance computing market, but rather for mass-market machines. The speculation is that a massively-multicore ARM array would be needed just to to get in striking range of multicore Intel chips.
By the way, I love the braggadocio of Smooth-Stone's CEO Barry Evans, who avers that the technology will "completely remove power as an issue in the data center"!
Stacey also reports today that Microsoft may be testing ARM-equipped servers, paired with solid state storage for use in their data centers that support online (cloud) services.
Now I try to distill these developments into layman's terms, if only to frame the follow-on questions. Here's my limited understanding, and what I'd like to know more about:
I understand that ARM chips are at least nominally more energy efficient that x86 architecture, because they have been optimized for mobile applications. When you're relying on a battery most if not all of the time, you write tight code and develop chips that sip rather than swig.
On the other hand, ARM chips simply can't do what x86 chips can in terms of outright performance, and whole swaths of the software market is written for x86.
My question: how does ARM architecture claim efficiency gains over x86? Is it because the chips are optimized for energy efficiency, or is there an inherent advantage in massively-parallel systems, where perhaps your idle- and low-usage states draw far less power than x86 chipsets that are just ticking over?
If the latter is true, then I would expect the bright minds at Intel and their partners will develop multicore machines that have much lower idle- and low-usage power draw, obviating the ARM advantage, if there is one.
If I learn ore from my contacts in the industry, I'll post further...
Frankly this is way old news. James Hamilton has been promoting this in multiple blog posts (4 to be exact) since January 2009. It is an interesting concept, but only works for certain workloads, mostly homogeneous. This may work well for the Google and Microsofts of the world, but I'd like to see it crank through an ERP, CRM, or batch processing system (can you imagine a batch process on an ARM core? not today anyway, it'd take forever!)
http://perspectives.mvdirona.com/2009/01/15/TheCaseForLowCostLowPowerServers.aspx
Posted by: Joe Polastre | 04/16/2010 at 07:12 PM