A Brain in a Band-Aid Box

A recent story caught my attention. It was about the new Dell XS11-VX8. A few things took me aback. First, it is not an Intel or AMD platform, and second, it is not multi-core. To be more specific it uses a VIA Nano processor (The Nano is similar to the Intel Atom processor, which are found in many low power netbook designs.) Now the interesting part, Dell has designed server modules each of which is about the size of a 3.5-inch disk storage bay (slightly longer). The module is a x86 64-bit Nano server that can hold up to 3 GB of main memory, an optional 2.5-inch disk or SD card, and provides two GigE ports. Dell reports an idle power draw of around 15 watts that can increases to peak of 29 watts under load. Twelve of these can be housed in a 2U chassis. They are intended for China and India where power is at a premium.

How “un-Dell.” Of course, it is reported to be customer driven and not a widely available product, but I suspect it will garner some interest in many areas. The choice for the Nano is interesting as well. Evidently, the initial customers wanted virtualization and the Intel Atom does not support the Intel VT instructions needed for Xen or KVM. The cost for a fully populated 12 processor (one core each) is stated to be about $4800 or about $400 per server node. This is not a “blade solution”, but the servers nodes share power and cooling from a 2U chassis.

How “Un-multicore.” We live in the age of multi-core. So why build a bunch of low power single core nodes? First, to be clear, VIA is not immune from the multi-core trend as it will be sampling a dual core Nano version later this year. In any case, this product seems to try and fit into a price-performance-power envelope that is not multi-core friendly. Let’s consider power. Each unit is said to use less than 30 Watts peak. Each 2U chassis then requires about 360 Watts. This power requirement is about half as much as you would need for an 8 core Nehalem solution. For the same power budget you could probably run 24 Nano servers vs 8 Nehalem cores (or one Nehalem server). Of course this is all back of the envelope ballpark numbers, but that means the Nehalem cores must be 3 times faster than the Nano cores to get parity in computational performance. Don’t worry, I bet the Nehalem’s are faster, but a power-performance analysis brings the Nano a into the same ball park as the Nehalem.

While we are at it, lets talk about memory bandwidth. Again, I don’t have hard numbers, but with some HPC codes memory bandwidth is important. While the Nano may have slower memory than the Nehalem, we have to account for the shared nature of multi-core node. Any time you share memory, you open yourself to possibility that bandwidth may not be consistent. I discussed this at length previously. In essence, your performance depends on how you share the a multi-core processor with other codes. With a single core processor and a private bank of memory, you don’t need to worry about this issue. One core with lots of private memory is nice ™.

To be clear, I am not advocating running out and buying Via Nano processors to do clustering. Indeed, there is more to it than simple comparisons. Although, I’ll bet some of the Monte Carlo guys are licking their chops at the Nano hardware based Random Number Generator that can produce up to 12 million random numbers per second. In a sense, the more low power/low speed processors you use the better your application scalability must be. And, nothing kills scalability like a poor interconnect. If want to see how low power HPC is done right, visit SiCortex. (Alas, SiCortex has ceased operations)

The point I wish to make is the added dimension of power and cooling may change a few things. Many customers now consider price-performance-power vs just price-performance. Indeed, SuperMicro has introduced an Atom based server as well. So much for netbook chips being used only for netbooks. If you are interested in Nano vs Atom Comparisons check out Intel Atom vs.VIA Nano and Netbook platforms: VIA Nano v. Intel Atom (The later includes some data for Intel E5200 as well so you can get a feel for the overall performance of the Nano and the Atom.)

And now, a true confession, in the movie 2001: A Space Odyssey when David Bowman goes into HAL 9000′s “brain” and starts removing small functional parts, I thought what a nice design. A brain in a band-aid box kind of approach. Of course the movie made it look really cool with the special effects, but, the idea of building a computer from small plug in parts seems almost natural. Small low power plugable nodes, now what could one do with a pile of those?