SGI is also launching a new product today. The launch is in DC, so if you’re hanging about the Nation’s capital with an itch to learn more about HPC, skip the monuments and head over to the launch.
ICE is SGI’s second generation blade solution. It’s Xeon-based (set for dual- or quad-cores) and is configured more like a traditional cluster than SGI’s previous offerings; i.e., no shared memory on this one. Where the 4700 is Itanium-based and will live on to aim at the very high end of the market, ICE is aimed at enterprises and departmental solutions where power, cooling, and manageability are drivers.
The system looks really interesting. You can get 512 cores in 128 nodes (on quad-core Xeon chips) in a single rack. A rack is cable-less up to 128 nodes (all the connections are by a passive backplane), which should smooth installation and management.
Nodes are interconnected by on-board DDR InfiniBand configured as a torus; 4x at node, and the switch blades can be configured with either 4x or 12x. Putting the InfiniBand on board should improve reliability, eliminate concerns about PCI saturation, and also saves on power and space in the configuration.
The energy picture is also interesting; one fully populated rack will draw about 45W/core, or about 23KW. The 4700 draws about 65W/core, so this is moving in the right direction.
The power picture is good, but I’m especially interested in this as a diversification of SGI’s product offering.
The HPC community needs shared memory, and right now that means it needs SGI. The trouble for SGI’s balance sheet is that we don’t need HPC in every procurement. This means that for the bulk of procurements not driven by a need for shared memory SGI’s products have been at a cost disadvantage and customers have been unwilling to buy technology they didn’t need. Now, however, SGI has a solution that builds on the engineering investments it’s already made with the Altix and Origin lines, but allows it to compete apples-to-apples with products from the other non-shared memory vendors.