Bringing optics inside the box: recent progress and future trends

Abstract

Needs and requirements for optical interconnects in next generation servers are outlied. Related results on equalizing and characterizing high speed multimode links, building and testing parallel I2xlOGbit/sec transceivers, and highly parallel silicon photodetectors are presented. 2 4 INTERCONNECT NEEDS AND REQUIREMENTSBecause of their ability to transport data at high speeds over long distances, optical fibers move the bulk of information in backbone networks. Because of the high cost of installing and maintaining the fibers in the gmund, high cost transceivers are acceptable, and single mode fibers (SMFs) and highly wavelength stable lasers are used.Since the mid-l9Ws, optics has also made increasing inroads into the LAN and SAN environments. In these environments, standards such as Ethemet, Fibre Channel, and Infiniband predominate. Here multimode fiber (MMF) l i are mainly used because the cost of the transceivers is much lower than single mode. There is a large base of installed MMF in buildings and on campuses, and there is a desire to continue to use that fiber but at higher aggregate data rates. The distance for such premises wiring is up to 550 m, but few existing MMF /transceiver combinations support this distance at 10GbitIsec. The cost of Dense WDM transceivers is generally prohibitive for these applications, and DWDM requires SMFs. Higher aggregate bandwidth could be achieved using Coarse WDM on MMF fihers, for which standards exist, but products have not been widely offered by manufacturers or accepted by users for a variety of technical and cost reasons.The above uses for optics are "outside the box". At the present time, server and switch-router designers are seriously looking at using optics "inside the box" to enable higher connectivity within the system. In general, the needed communication bandwidth inside high-end systems is increasing about 3X every 2 years. The bandwidth increases faster tban other performance metrics in the system, such as clock speed and number of microprocessors (both of which are increasing at about 2X every 2 years) as the needed bandwidth is the product of these other m h c s , with some relief coming from larger caches and better system software design. By 2010, high end servers are expected to require -4OTerabitslsec of interconnection between internal nodes, where a node is a microprocessor and its associated cache. Presently, copper interconnects are used in these systems, but optics should gradually replace copper cables (initially) and then circuit hoard traces (later) in many of the interconnects, provided the technology and cost are acceptable. The primary advantages of optics over copper come from the higher bitrate per l i and the higher density of links for optics. The cost of optics, at 10-IOOX that of copper interconnects is a major roadblock. The 1OX figure is for fiber optics compared to electrical cables at -1Ometer lenghts, and the lOOX figure is fiber compared to copper circuit board traces and backplane connectors. Additional issues slowing the acce...

Reconfigurable and adaptive photonic networks for high-performance computing systems

“…Optical interconnects offer several well-known advantages for HPC systems such as higher spatial and temporal bandwidths, lower cross talk independent of data rates, higher interconnect densities, better signal integrity at high frequencies, lower signal attenuation, and lower power requirements at higher bit rates [2,3,[10][11][12][13][14], all of which could potentially achieve the much desired high bit rates data communication at a much reduced power level at the boardto-board distances of 0.1-1 m.…”

“…It is widely accepted that Moore's law [and the more recent International Technology Roadmap for Semiconductors (ITRS)] growth rate in available transistors will continue for at least the next decade, thereby strengthening further growth of HPC systems. Near-term projections call for HPC systems with computing power in the hundreds of teraflops, off-chip bandwidth in the range of 4-20 Tbits/s and aggregate interprocessor communication bandwidth or network bandwidth around 40 Tbits/s [2][3][4]. However, with the ITRS projects, although per-chip performance continues to improve at a rate of approximately four times, the total off-chip input/ output (I/O) bandwidth (pin count times the bit rate per pin) will increase by approximately 2.7 times [4].…”

“…However, with the ITRS projects, although per-chip performance continues to improve at a rate of approximately four times, the total off-chip input/ output (I/O) bandwidth (pin count times the bit rate per pin) will increase by approximately 2.7 times [4]. As clock rates increase to the multigigahertz range, and this difference in improvement rates continue, electrical signaling and interconnect problems, such as skin effect, cross talk, electromagnetic interference (EMI), dielectric imperfections, attenuation, clock skew, and power dissipation are predicted to become the ultimate bottlenecks for HPC systems both at the chip-to-chip and board-to-board levels [2,3,[5][6][7]. The major bottleneck is the limited bandwidth at higher bit rates and longer communication distances.…”

Performance adaptive power-aware reconfigurable optical interconnects for high-performance computing (HPC) systems

“…The static routing and wavelength allocation (RWA) for inter-board communication for a R (1,4,4) system is shown in Figure 1. For inter-board communication, different wavelengths from various boards are selectively merged to separate channels to provide high connectivity.…”

“…The increasing bandwidth demands at higher bit rates and longer communication distances in high-performance computing (HPC) systems are constraining the performance of electrical interconnects [1,2,3,4,5,6]. This has given rise to opto-electronic networks that can support greater bandwidth through a combination of efficient multiplexing techniques (wavelength-division, time-division, and spacedivision) for board-to-board and rack-to-rack interconnects.…”