For most customers, a cluster built from 1U’s or blades will be sufficient. However, for the really large computing systems (particularly for technical computing customers), the cost to maintain a cluster may outweigh any savings from using commodity components. For these customers, a handful of vendors supply real supercomputers.
Perhaps the most famous supercomputer company is Cray, which has gone through a number of different owners over the years and thus barely resembles its original carnation. None the less, Cray still produces a line of widely regarded machines. The XD1 is a Linux-on-Opteron computer that features Cray’s proprietary RapidArray interconnect, a switched fabric. The XT3 is another Opteron machine, but uses Cray’s UNICOS/lc operating system and features the SeaStar-based torus network. The X1E is a ccNUMA that uses UNICOS/mp and Cray’s own vector processor.
A more recent entry into supercomputing is SGI, which until recently was better known for its visualization products. The flagship Altix line features Linux-on-Itanium with the NUMAlink interconnect. As the network name implies, Altix is a ccNUMA.
Of course the perennial IBM has the Blue Gene as its latest entry into supercomputing. This machine features PowerPC processors and a proprietary light-weight kernel for the compute nodes. It also has both a torus network and a tree network, the former for point-to-point communication and the later for collective communication.
NEC, of Earth Simulator fame, boasts the SX-8 series, which uses vector processors, the SUPER-UX operating system, and the IXS switch-based interconnect. The ccNUMA TX7 series uses Itanium and supports either Linux or HP-UX.
Startup Linux Networx (LNXI) aims to have customizable systems based on commodity components. To that end, the LS-1 series features Linux on Operton with the customer’s choice of Ethernet, Myrinet, or InfiniBand. The LS/X is also Linux on Opteron, but with the InfiniPath interconnect.
The offerings listed here provide some insights into the direction of the highest end of technical computing. While a few vendors such as Streamline Computing choose to customize every system from scratch, most integrators prefer having a set product line. This strategy makes sense as it ensures that the integrator can test all possible components to discover the exact permutation that works best. Indeed, this is the strategy used by most PC manufactures.
At first glance, it would appear vendors tend to sport similar components. Most operating systems are either Linux or Unix and most processors are either Opteron or Itanium. Most networks, however, do tend to be proprietary. Also, the prospect of shared memory programming via ccNUMA is present in some cases.
A quick glance at the market, however, will reveal different trends. Firstly, most technical software is developed as distributed memory, so ccNUMA does not appear to be a big selling point for large computers. Indeed, Cray and SGI have both recently faced financial difficulties, whereas Linux Networx and IBM’s HPC unit are doing quite well. NEC has a huge lead in Japan and a respectable presence in Europe, but has not had nearly the same success in North America. It would appear then that NEC’s market penetration has more to do with its geography (its the only Asian vendor listed here) rather than technical issues.
IBM and LNXI both represent two extremes in the marketplace, using either entirely proprietary components or entirely commodity components. It would appear that each is “exciting the base” rather than aiming for a middle ground. Pundits in the enterprise computing market indeed point to the success of Dell (low price) and IBM (high service) in contrast to the stumble of HP. (Though to be fair, HP also bought Compaq whereas IBM sold off its PC unit to Lenovo; this strategy in the PC market may have as much to do with success in the enterprise market, given the allocation of corporate resources, etc.)
Cray and SGI are both attempting to fight back by consolidating their respective product lines and playing to “adaptive” or “reconfigurable” supercomputing. Both the Altix and the XD1 feature Xilinx Virtex FPGAs, thereby allowing the end user to create a co-processor for common application-specific tasks. Only time will tell whether this will be enough.