SGI has announced the introduction of Prism XL — a new product for accelerator-based high-performance computing. The device was designed to fully leverage the scale and speed of accelerators, enabling users to scale to the tens of petaflops or large teraflop range.
The Prism XL boasts extreme density. Whereas x86 processors may require up to 100 cabinets to deliver a petaflop of computing performance, SGI Prism XL delivers similar performance and scale in a single cabinet, and a 20-petaflop double precision deployment can be achieved in as few as 100 cabinets. This represents a 100:1 reduction in facility floor space and up to a 30 percent reduction in overall total cost of ownership.
Also announced was the SGI STIX architecture, the basic building block for the Prism XL. STIX architecture moves beyond the blade to a “stick” form factor while also accelerating the pace of computing toward the exascale range.
Based on a PCIe infrastructure for accelerator deployment while being accelerator-agnostic, STIX supports the latest accelerator cards from NVIDIA, AMD and Tilera, and up to a 300W PCIe card. A ‘stick’ is used as the integral unit, with each stick containing two “slices” of a full-length, full-height PCIe gen 2 x 16 slot and a single-socket AMD 4100-powered motherboard.