Jump to content

NUMAlink

From Alfawiki
NUMAlink
Company typeTechnology
IndustryComputer hardware, Interconnect technology
PredecessorStanford DASH interconnect
Founded1990
FounderSilicon Graphics (SGI)
Area served
Distributed shared memory ccNUMA architectures
ProductsNUMAlink interconnect technology
OwnerSilicon Graphics (SGI), later Hewlett Packard Enterprise (HPE) partnership

NUMAlink represents a system interconnect technology engineered by Silicon Graphics (SGI) specifically for its distributed shared memory ccNUMA architectures. Initially developed for the Origin 2000 and Onyx2 platforms, it was marketed as "CrayLink" during the period when SGI briefly owned Cray Research.

An original equipment manufacturer (OEM) partnership was established between Hewlett Packard Enterprise (HPE) and Silicon Graphics International (SGI) to incorporate NUMAlink technology into select mission-critical server solutions.

Generation Launch Year Bandwidth[lower-roman 1] Associated Systems
Stanford DASH circa 1990 0.12 GB/s[lower-roman 2]
NUMAlink 2 1996 0.8 GB/s Origin 200, Origin 2000, Onyx2
NUMAlink 3 2000 1.6 GB/s Origin 3000, Altix 3000
NUMAlink 4 2004 3.2 GB/s Altix 4000
NUMAlink 5 2009 7.5 GB/s Altix UV
NUMAlink 6 2012 6.7 GB/s UV 2000, UV 3000, UV 30
NUMAlink 7 2014 14.9 GB/s UV 30EX, UV 300 / HPE Integrity MC990 X
NUMAlink 8 2017 13.3 GB/s (HPE) Superdome Flex, Compute Scale-up
Notes

NUMAlink 2

SGI did not designate any product as NUMAlink 1, considering the interconnect used in the Stanford DASH project as the initial generation. NUMAlink 2, also known under the CrayLink brand, was introduced in October 1996 and deployed in the SGI Onyx2, SGI Origin 200, and SGI Origin 2000 platforms. This generation employed the Hub ASIC interface and achieved a maximum bandwidth of 1.6 GB/s through two separate unidirectional PECL links operating at 400 MHz with a 16-bit width, each providing 800 MB/s.

NUMAlink 3

Released in 2000, NUMAlink 3 served as the interconnect for the Origin 3000 and Altix 3000 systems. It doubled peak throughput to 3.2 GB/s by utilizing two unidirectional links each capable of transferring data at 1.6 GB/s. This period also saw the introduction of a modular architecture strategy termed NUMAflex.

NUMAlink 4

Introduced in 2004 alongside the Altix 4000, NUMAlink 4 further increased peak throughput to 6.4 GB/s by employing two unidirectional channels each delivering up to 3.2 GB/s.

NUMAlink 5

Debuting in 2009 for the Altix UV series, NUMAlink 5 enhanced peak bandwidth to a total of 15 GB/s via two unidirectional links rated at 7.5 GB/s each, meeting the demands of high-performance computing environments.

NUMAlink 6

The sixth iteration, introduced in 2012, was integrated into the SGI UV product line including models UV 2000, UV 3000, and UV 30. It provided a bidirectional peak bandwidth of 6.7 GB/s and supported scalable configurations up to 256 sockets featuring coherent shared memory capacities reaching as high as 64TB.

NUMAlink 7

Launched in 2014, NUMAlink 7 powered systems such as HPE Integrity MC990 X along with several SGI UV variants including UV 300, UV 30EX, UV 300H, and UV 300RL. This generation offered nearly 15 GB/s bidirectional peak bandwidth and supported architectures with up to 64 sockets and coherent shared memory totaling up to 64TB.

NUMAlink 8 (Flex ASIC)

Launched in 2017, NUMAlink 8 represents the eighth iteration of the interconnect technology utilized in HPE Superdome Flex and Compute Scale-up platforms. It offers a bandwidth of 13.3 GB/s per port, facilitating a maximum bisection peak bandwidth of 853.33 GB/s within a 32-socket configuration that supports up to 48 TB of coherent shared memory. This cutting-edge interconnect is designed to optimize scalability and data transfer rates, catering to demanding enterprise and high-performance computing (HPC) applications in expansive computing infrastructures.

See also

References

External links


Cite error: <ref> tags exist for a group named "lower-roman", but no corresponding <references group="lower-roman"/> tag was found