The new SGI is continuing the supercomputing legacy of its namesake processor.
At the Supercomputing 2009 show Nov. 16, SGI unveiled what officials are calling the fastest and most scalable shared-memory system.
Powered by Intel's upcoming "Nehalem EX" Xeon processors, the Altix UV can scale from 32 to 2,048 processing cores and can support up to 16 terabytes of shared memory in a single image, all within four server racks.
"It just looks like all the memory is in a single addressable pool," Geoffrey Noer, senior director of product marketing at SGI, said in an interview.
The bladed system-once code-named "UltraViolet"-can run at 18.6 teraflops, or 1 trillion floating point instructions per second.
Aimed at HPC (high-performance computing) environments, and large-scale database and data analytics workloads, the Altix UV also takes advantage of SGI's high-speed 15GB-per-second NUMAlink 5 interconnect and MOE (MPI Offload Engine) acceleration technology.
With such features as the ability to use four-, six- or eight-core Xeon chips, scalability is a key attribute for the new system, according to Noer.
"You can really start small and scale over time," he said.
The Altix UV can reach a petaflop of performance by using the NUMAlink 5 interconnect to link the four racks-all 256 sockets-to eventually scale up to 32 racks and 16,384 sockets.
The use of industry-standard technology also enables the Altix UV to run unmodified SUSE and Red Hat Linux operating systems.
The system comes with the SGI ProPack and ISLE software for management and deployment used with other Altix systems.
"It's all familiar to SGI customers ... but it's also a comprehensive stack for people new to SGI," Noer said.
Altix UV will come in two versions. The Altix UV 1000 is made for maximum scalability, shipping as a fully integrated cabinet solution with up to 256 sockets (2,048 cores) and 16TB of shared memory in four racks, with performance of up to 18.6 teraflops.
The midrange Altix UV 100 comes in a rack-mount 3U form factor. The system will scale to 96 sockets (768 cores) and 6TB of shared memory in two racks for up to 7 teraflops of performance.