Editing Silicon Graphics (section)

==High-end server market==
SGI continued to enhance its line of servers (including some [[supercomputer]]s) based on the [[SN architecture]]. SN, for Scalable Node, is a technology developed by SGI in the mid-1990s that uses [[Non-Uniform Memory Access#Cache coherent NUMA (ccNUMA)|cache-coherent non-uniform memory access]] (cc-NUMA). In an SN system, processors, memory, and a bus- and memory-controller are coupled together into an entity called a node, usually on a single [[circuit board]]. Nodes are connected by a high-speed interconnect called [[NUMAlink]] (originally marketed as [[CrayLink]]). There is no [[internal bus]], and instead access between processors, memory, and [[I/O]] devices is done through a [[switched fabric]] of links and [[router (computing)|router]]s.

Thanks to the cache coherence of the [[distributed shared memory]], SN systems scale along several axes at once: as CPU count increases, so does memory capacity, I/O capacity, and system [[bisection bandwidth]]. This allows the combined memory of all the nodes to be accessed under a single [[Operating system|OS]] image using standard [[shared-memory synchronization]] methods. This makes an SN system far easier to program and able to achieve higher sustained-to-peak performance than non-cache-coherent systems like conventional [[cluster computing|clusters]] or [[massively parallel computer]]s which require applications code to be written (or re-written) to do explicit [[message-passing]] communication between their nodes.

The first SN system, known as SN-0, was released in 1996 under the product name [[SGI Origin 2000|Origin 2000]]. Based on the MIPS [[R10000]] processor, it scaled from 2 to 128 processors and a smaller version, the [[SGI Origin 200|Origin 200]] (SN-00), scaled from 1 to 4. Later enhancements enabled systems of as large as 512 processors.

The second generation system, originally called SN-1 but later SN-MIPS, was released in July 2000, as the [[Origin 3000]]. It scaled from 4 to 512 processors, and 1,024-processor configurations were delivered by special order to some customers. A smaller, less scalable implementation followed, called Origin 300.

In November 2002, SGI announced a repackaging of its SN system, under the name Origin 3900. It quadrupled the processor area density of the SN-MIPS system, from 32 up to 128 processors per rack while moving to a "[[fat tree]]" interconnect topology.

In January 2003, SGI announced a variant of the SN platform called the [[Altix|Altix 3000]] (internally called SN-IA). It used [[Intel]] [[Itanium]] 2 processors and ran the [[Linux]] operating system kernel. At the time it was released, it was the world's most scalable Linux-based computer, supporting up to 64 processors in a single system node.<ref>[http://www.linuxjournal.com/article/6440 Scaling Linux to New Heights: the SGI Altix 3000 System] Linux Journal, January 2003</ref> Nodes could be connected using the same [[NUMAlink]] technology to form what SGI predictably termed "superclusters".

In February 2004, SGI announced general support for 128 processor nodes to be followed by 256 and 512 processor versions that year.

In April 2004, SGI announced the sale of its Alias software business for approximately $57 million.<ref>{{cite press release
  |url = http://www.sgi.com/newsroom/press_releases/2004/april/alias.html
  |archive-url = https://web.archive.org/web/20071205072024/http://www.sgi.com/company_info/newsroom/press_releases/2004/april/alias.html
  |archive-date = December 5, 2007
  |title = Silicon Graphics Sells Alias Software Business
  |access-date = March 29, 2011
  |date = April 15, 2004
  |publisher = Silicon Graphics
}}</ref>

In October 2004, SGI built the supercomputer [[Columbia (supercomputer)|Columbia]], which broke the world record for computer speed, for the [[NASA Ames]] Research Center. It was a cluster of 20 [[SGI Altix|Altix supercomputers]] each with 512 Intel Itanium 2 processors running Linux, and achieved sustained speed of 42.7 [[1000000000000 (number)|trillion]] floating-point operations per second ([[FLOPS|teraflops]]), easily topping [[Japan]]'s famed [[Earth Simulator]]'s record of 35.86 teraflops. (A week later, [[IBM]]'s upgraded [[Blue Gene]]/L clocked in at 70.7 teraflops.)

In July 2006, SGI announced an SGI Altix 4700 system with 1,024 processors and 4 [[terabyte|TB]] of memory running a single Linux system image.<ref>{{cite press release
 |url         = http://www.sgi.com/company_info/newsroom/press_releases/2006/july/stream_1024p.html
 |archive-url  = https://web.archive.org/web/20110607111644/http://www.sgi.com/company_info/newsroom/press_releases/2006/july/stream_1024p.html
 |archive-date = June 7, 2011
 |title       = SGI Altix Again Crushes World Record for Memory Bandwidth
 |access-date  = March 29, 2011
 |date        = July 17, 2006
 |publisher   = Silicon Graphics
 |url-status     = dead
}}</ref>