Large shared-memory computing seems to be making a bit of a comeback. The two players we have already talked about ScaleMP and 3Leaf, use primarily off-the-shelf servers (with a special ASIC, in 3Leaf’s case) along with Infiniband connections and specialized software to build cache-coherent, shared-everything systems.
With this, they can build systems spanning 16 motherboards – which is pretty damned big when you factor in the core counts you can get today.
SGI is taking a different, and much larger, swing at this with its new UltraViolet Nehalem-based clusters. According to Geoff Noer of SGI, these boxes will scale to 2,048 cores and 16TB of memory with a single operating system instance – unmodified SUSE or Red Hat Linux.
Our buddy Timothy Prickett Morgan wrote about these systems in great detail and did a great job of laying out the rationale behind these boxes and the numbers on their feeds and speeds.
We saw the system in its skins at the SC09 show and had a quick chat with Geoff Noer to get the specifics…click here to look at the video…