This week Intel announced its intention to deliver chips for the burgeoning microserver market. Microservers are envisioned as the next big thing in hyperscale datacenters that are delivering Web content and lightweight software apps for the masses.

That market is being carved out by vendors like SeaMicro, Marvell, Dell, Tilera, Calxeda, and some others, who envision a growing opportunity for datacenter workloads that don't need energy-sucking Xeons or Opterons to do their job. A handful of these enterprising companies have developed their first-generation microservers, mostly based on inexpensive, power-sipping Atom and ARM processors. In fact, because of the scale of these Internet datacenters, energy efficiency, price-performance, and density have become the prime considerations for building Web-based infrastructure.

Intel sees this market as comprising no more than 10 percent of overall server sales over the next four or five years, but that is large enough to get the chipmaker's attention. The company's announcement this week revealed a roadmap that includes both Xeon and Atom x86 processors aimed at the microserver space. That includes its recently released E3-1260L and E3-1220L Xeon chips, rated at 45 and 20 watts, respectively. In the second half of 2011, the company plans to deliver a 15-watt Sandy Bridge server part for this market. The first Atom server chip doesn't arrive until 2012, and Intel is spec'ing that chip to come in at somewhere under 10 watts.

That didn't stop SeaMicro from jumping the gun and building microservers with the just-released Atom N570, a 64-bit dual-core chip that has a TDP of just 8.5 watts. The new gear, announced last month, is an upgrade from the original SeaMicro server the company introduced in 2010, which used 32-bit single-core Atom N530 processors. Now SeaMicro has latched on to the up-market N570 (officially targeted for netbooks and notebooks) and is putting 256 of them in a 10U box. The CPUs are connected via SeaMicro's proprietary fabric that delivers up to 1.28 terabits/second of aggregate bandwidth. In all, a nifty little 512-core x86 server that lists for $148,000.

So what does all this have to do with high performance computing? Right now, not a whole lot. From a performance point of view these x86 microserver chips are pretty wimpy compared to their more muscular Xeon brethren. For example, the quad-core E3-1260L runs at just 2.4 GHz compared to a traditional HPC-worthy Xeon, like the quad-core 3.2 GHz X5672. The E3-1260L also has only two memory channels versus three for the X5672, and only 8 MB of cache versus 12 MB for the X5672.

Basically you get about two-thirds of an HPC chip that uses about half the power. That might sound like a good trade-off until you consider that the E3's are designed only for single-socket machines, so you would have to buy twice as many servers to get the same number of cores as in a traditional dual-socket box.

The starker contrast is the Atom N570 being used in the new SeaMicro gear. That processor runs at just 1.6 GHz, sports only 1 MB of cache, and can support a maximum of 4 GB of memory. (Oh, and as far as I can tell, there's no support for ECC memory in any of the current Atom chips.) With half the number of cores running at half the speed, and with much less cache and memory, N570 performance will be just a fraction of the X5672's.

On the other hand, at 8.5 watts, the power draw on the N570 is less than a tenth that of the X5672. And here's the real kicker: the Atom is just $86 in quantities of a thousand. That's about 1/16 the $1,440 price listed for the Xeon part.

The future Atom silicon that Intel will be officially targeting for microservers is likely to be more performant than the current N570 (and presumably include ECC as well). But the chipmaker has to thread a needle here. It can't sell Xeons at $1,000-plus a pop if it's also offering $100 Atoms that are just a few time slower. The ARM makers, by the way, will have no such conflict.

Intel imagines there's some continuum of workloads, where 90 percent of users will stick with the faster CPUs represented by the traditional Xeon parts, and the rest will want stripped-down Xeons or pumped-up Atoms. The rationale is that many enterprise codes still rely on single-core performance and can't be parallelized into a gazillion threads that take advantage of voluminous low-performing cores.

But many Web serving applications, like Facebook or Google search, naturally decompose into multiple threads that run more or less independently from one another, and without any messy virtualization. Also, these codes tend not to do a great deal of number crunching, which is just fine for Atom and ARM chips since they are not particular adept at executing more complex operations like floating point instructions.

That's not to say these first-generation microservers are worthless for high performance computing. Where codes are easily decomposed into embarrassingly parallel execution and don't rely on lots of floating point performance (like genetic sequencing apps), this simplified architecture could find great utility.

The truth is, though, no one really knows how the microserver business will play out or even what the size of the market will be. Web applications are evolving rapidly and may end up needing more powerful processors than recycled ARM or Atom designs. But Intel's entry into the market marks something of a turning point. When the biggest chipmaker in the world decides to go after a smaller, lower margin market, we should assume it has done so for a very good reason.