Wimpy cores versus brawny cores

Calxeda and Tilera are two ARM-based products that offer are a “wimpy” core approach.  For the same level of work performed, a large number of less powerful processor cores will use less power than a smaller number of more powerful processor cores.  Since power is a huge cost in running data centers, less power consumed may translate into lower overall costs.

However, the wimpy core approach is unsuitable for many applications.  The Google paper “Brawny cores still beat wimpy cores, most of the time” explains why.  The short story is that many problems map poorly to wimpy cores. The corollary of the paper’s title suggests that there are niche applications where the wimpy core approach is the best.

Intel has a wimpy core solution with its Atom and Centerton processors.  The ARM-based solutions take the wimpy core approach to an extreme compared to Intel.  Their products are far “wimpier” than Intel’s products and may do well in a niche within the wimpy core niche.  For a very large number of extremely wimpy cores, the ARM-based solutions may have an advantage over x86 cores as each x86 core will have some overhead related to the x86 instruction set.  Within this extremely small niche, I am not sure if Calxeda and Tilera have had much success in gaining traction in the server market.

ARM-based server products currently have low rates of adoption.  Facebook may or may not choose ARM in the future: “Facebook is evaluating ARM’s chips for their low-power characteristics and reasonable compute capability. However, it won’t be until the launch of 64-bit variants next year that it gets serious about the processors, the engineers indicated.” (http://www.zdnet.com/facebook-stretches-arm-chips-in-datacentre-tests-7000004707/)  HP’s Project Moonshot was originally designed with ARM-based processors in mind but will first ship with Intel’s Centerton.

Seamicro

Seamicro is a startup company (purchased by AMD) that produces an innovative solution for the server market.  In a typical server, each CPU needs to be supported by a motherboard with various chips on it (e.g. networking, storage) and at least one hard drive.  Seamicro’s solution allows multiple CPUs to share all these resources.  This produces savings on hardware costs as you need a fraction of the networking and storage chips for each CPU and a fraction of a hard drive.  Power costs also go down since a lot less hardware is being powered up.

Seamicro’s technology plays well with the wimpy core approach.  One problem with the wimpy core approach is that less powerful processors means that you need more servers… leading to higher hardware costs.  Seamicro’s approach reduces hardware costs and shifts the economics more in favour of wimpy cores.

AMD and Seamicro will likely pose far more competition to Intel than ARM-based servers.  I expect that AMD will design processors that play well with Seamicro’s product.  Currently there is a Seamicro product that uses Intel Xeon CPUs with many features on the Xeon CPU turned off (e.g. SATA, integrated graphics, etc.).  The obvious evolution is for AMD to produce a similar chip without these features.

Semiaccurate has a great article on AMD’s purchase of Seamicro.  While the author portrays AMD’s purchase as a huge loss for Intel, his predictions about the future haven’t always been accurate in the past (e.g. Facebook using ARM servers; the “tectonic shift” didn’t really materialize).

Intel Centerton

Intel’s approach to the microserver market is to basically put an entire server on a chip.  The current trend in all areas of computing has been to do a “system on a chip”.  Instead of using many different pieces of silicon, manufacturers have been merging everything together onto a single piece of silicon.  This saves on packaging costs and saves space in smartphones.  Whereas Seamicro has each CPU sharing networking chips (with a chip facilitating the sharing of resources), Intel simply combines the CPU and networking chip on a single piece of silicon.

I couldn’t find much data on the relative costs and performance between Centerton’s approach and Seamicro’s approach.

Where the future is headed

Intel sees microservers growing but it expects them to take up 10% of the overall server market by 2015 from 1% to 2% today.  From Intel’s latest 10-K, the data center group had revenues of $10,129 so microservers might only be a $1B/yr market for Intel.

I think that ARM-based products will not pose a serious competitive threat to Intel.  Even if they are successful, these products are only currently targeting a niche within the microserver niche.

In the overall server market, RISC-based products (e.g. Power, Sparc) will likely continue their trend towards lower market share.  Intel’s Itanium CPU, which does not use the x86 instruction set, will lose share to Intel’s x86 Xeons over time as Xeons are just as fast and will offer the same high-reliability features.  AMD will likely be Intel’s most dangerous competitor as it has high volume and an x86 license.