Optical Interconnect Opportunities for Future Server Memory Systems

Katayama, Y.; Okazaki, Atsuya

doi:10.1109/hpca.2007.346184

Cited by 21 publications

(16 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Relocation of the memory devices to be physically distant from the processor also frees up board space near the processor itself without significant impact on system performance [12]. This will give system designers more flexibility in designing board layouts: remaining board components, such as on-board routers or caches, can be placed closer to the processor or increased in number.…”

Section: Optically Connected Memorymentioning

confidence: 99%

“…The work in [12] explores the impact of accessing large banks of remote memory across optical links. The authors focus on symmetric multiprocessing (SMP) systems, and demonstrate that, within large SMP systems, the improved bandwidth from optical links outweighs the increased time-of-flight latency incurred from moving optically attached memory physically away from processing elements.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Building Data Centers With Optically Connected Memory

Brunina¹,

Lai

Garg

et al. 2011

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

Abstract-Future data centers will require novel, scalable memory architectures capable of sustaining high bandwidths while still achieving low memory access latencies. Electronic interconnects cannot meet the challenges presented by the need for multi-terabit off-chip memory data paths. In this work, the electronic bus between main memory and its host processor is replaced with a circuit-switched optical interconnection network. We investigate the impact of our optically connected memory system on large-scale architectures and experimentally validate the protocol using field-programmable gate array based processor nodes and a custom-designed memory controller. The processor communicates all-optically with multiple synchronous dynamic random access memory nodes using 4×2.5-Gb/s wavelength-striped payloads, operating error free with bit-error rates less than 10 −12 .

show abstract

Section: Optically Connected Memorymentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Building Data Centers With Optically Connected Memory

Brunina¹,

Lai

Garg

et al. 2011

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

show abstract

“…Even with a less aggressive use of optics, one should at least expect that optics will support most of the off-chip communications, including on-card links [11] between processors and processor-to-memory links [12]. This increased use of optics can only occur if the cost and power of an optical link can be substantially…”

Section: Projections For Future Supercomputersmentioning

confidence: 99%

Optical Interconnects in Future Servers

Kash

Benner²,

Doany

et al. 2011

Optical Fiber Communication Conference/National Fiber Optic Engineers Conference 2011

View full text Add to dashboard Cite

Optical interconnects are common in today's petascale supercomputers, and will become pervasive at the exascale during this decade. Technologies that can meet the challenging technological and economic requirements for the exascale will be reviewed. Optics in Petascale SupercomputersIn 2008, Roadrunner[1], the first petaflop supercomputer (i.e., >10 15 floating point operations per second) was constructed. The large number of microprocessor cores in such petascale machines must be interconnected with a high capacity communications network to permit efficient computation. Higher interconnect bandwidth will generally result in more efficient use of the microprocessors in real calculations. Prior to 2005, this network used electrical interconnects in essentially all supercomputers. Communication bitrates increased beyond ~2Gb/s starting in about 2005. For distances greater than about 20m, electrical interconnects were impractical and optics began to be used for these longer rack-to-rack interconnects (e.g., ASCI Purple[2]). For cost reasons, however, the shorter rack-to-rack interconnect in this machine was still electrical. By 2008, communication bitrates had increased to 5Gb/s (InfiniBand DDR), and the practical reach of electrical interconnects was shorter. At the same time, the cost of optics had decreased. As a result, for Roadrunner, all node-to-node communication was optical. However, in these examples, all the intra-node communication was electrical. Also, the conversion to optics was at the edge of the node/rack, requiring an electrical interconnect from the chip through several levels of packaging before conversion to photons. For machines with peak performance beyond 10PF, optics will play a far more important role. For example, in the planned POWER7-IH systems [3], all the board-to-board interconnects, will be optical. A major change in the packaging tightly integrates the optics onto the chip-level package. Optical transmitter and receiver modules are placed on the same ceramic Hub (switch) Module [4,5] , with each channel using a 10Gb/s signaling rate and 8b10b coding. Projections for future supercomputersBecause microprocessor clock speeds are not rapidly increasing, supercomputing performance improvements come from larger numbers of processing cores operating in parallel. For example [7], the #1 machine on the top500 list from June, 2002 was the NEC Earth-Simulator with 5120 single core processor chips, as compared to Roadrunner which has some 97,920 cores in 12,240 Cell Broadband Engine® processor chips plus 6,562 dual-core AMD Opteron® chips [8]. This increasing core and chip count drives increasing communication bandwidth at higher bitrates.Trends for optical interconnects in future supercomputers can be extrapolated from past growth. Historically, computational power has grown approximately 10x every 4 years [7], and this growth is expected to continue. Affordable supercomputing requires that the cost and power consumption of supercomputers grow more slowly than computational power. ...

show abstract

“…We propose opticallyconnected memory systems, leveraging optics' power-efficient distance immunity at rack-to-rack computer scales, high-bandwidth density, and time-of-flight latencies [2,3]. Future optically-connected memory designs should also effectively support the redundancy and overprovisioning functionalities that current systems typically use to improve performance and reliability [4].…”

Section: Introductionmentioning

confidence: 99%