SGI releasing Kepler based rackable unit
May 15, 2012 – With the announcement of the newest generation of the Tesla GPU the Kepler architecture, which is available as a K10 and K20 model, SGI has created a GPU based computing node that ia available in a rackable form.
The core of the system is a dual CPU system featuring the Sandybridge Xeon X5-2600 chips, 8 DIMM slots, three PCIe Gen 3×16 double-width expansion slots in a 1U standard depth chassis. The unit can accommodate up to four 2.5 inch hard disk drives, and dual-port GigE networking controllers. The system is also available in a 2U chassis, where it can accommodate up to ten 2.5 inch hard drives, four internal PCI-E 3.0 x 16 double-width expansion slots and dual-port GigE controllers.
The unit is available as single units or as of the SGI GPU Starter Kit with NVIDIA® Tesla® K10 GPU Accelerator in a pre-configured bundle of 10 nodes. The 10 node kit uses Infiniband connections and supports a GigE switch block for interface. For those on the migration path, the units support the same GDDR5 memory as previous units that used the Fermi GPU card. A K20 based unit is planned for the future. The starter kit and the single units currently only come with administrator tools, and there is no SDK development platform as yet. The unit supports CUDA, OpenGL and OpenCL for the GPU programming and is optimized for 32bit Floating Point computation.
