SGI announces one petaflop per cabinet servers

New platform could offer significant performance improvements.

SGI has announced that it is able to fit one petaflop of compute power into a single cabinet thanks to the use of graphics chips.

The company had been using Nvidia's Tesla general purpose graphics processing units (GPGPU) to power part of its server line-up aimed at researchers and engineers.

SGI has now announced that it plans to offer similar GPGPUs from rival firm ATI and niche chip designer Tilera, and has left the door open for other GPGPU vendors, saying that further cards may be connected using the PCI-Express interconnect.

The combination of machines containing these GPGPUs will allow it to fit one petaflop, or a thousand trillion floating point calculations per second, into a single server cabinet, according to the firm's calculations.

Nvidia has been extremely successful in the high performance computing market, and a Tesla powered system is ranked second in the prestigious Top 500 supercomputer list.

Given the prodigious performance, SGI, like its rivals, is keen to offer the widest possible range of servers which make use of GPGPUs.

The move comes amid Nvidia's problems with its latest Tesla computation boards that use chips based on its Fermi architecture. The problems have led to Nvidia scaling back on performance and increased power utilisation and increased design work for vendors such as SGI.

SGI senior vice president and chief technology officer Eng Lim Goh said that his firm's Hybrid Computing Platform offered a "radically differentiated server technology" and that it would "enable users to attain new levels of scalability and speed".

SGI plan to ship products with its Hybrid Computing technology by the end of 2010.