Open Access
Add to BibSonomyAbstract
A performance comparison between the electrical Cu-based backplane and a full-optical fiber-based backplane is presented in terms of capacity and power consumption. By means of systematic simulations we find the electrical configuration, which allows to optimize the Cu-based backplane by exploiting the best technologies available today. On the other hand, a fiber-based optical backplane is proposed by exploiting the most performing VCSEL sources. Limitations of the electrical and optical approaches are discussed, considering their capabilities to support up to about 25-Gb/s transmission and the possibility to evolve towards higher bit-rates.
©2013 Optical Society of America
1. Introduction
The exponential growing of the requested bandwidth capacity of backplane component is driven by information technology and telecommunication equipment [1]. The solutions currently used are based on copper interconnections via backplane, driven by a suitable transceivers subsystem. Looking at high-end telco equipment transceivers are directly integrated on the same silicon chip used for the switching function or Input/Output (I/O) function (see Fig. 1). The transmitting and receiving interfaces, the connectors, and the Printed Board (PB) technology evolved and allowed us to increase the single interface bandwidth. In parallel the amplitudes of the transmitted signals has been reduced, to limit power consumption and increase the integration. In Fig. 2, as an example, we report the power consumption evolution of Optical Multi Service Network / Packet Transport Network (OMSN/PTN) product family exploited by Alcatel Lucent (ALU). The power figures take into account the complete chain from two blades attached to the backplane, including driver and receiver. According to the specific application, proprietary ASICs or standard commercial components are used (for ALTERA device Altera Stratix GT 28Gb device is considered where the Physical Media Attachment - PMA - includes serialization and de-serialization operations, pre-emphasis, Clock Data Recovery - CDR - and equalization). In general we can observe the progressive capacity to enhance the power density in a single device and while the consumption measured in mW per Gb/s has had an exponential decrease, the consumption of the interface including Serdes, CDR, pre-emphasis and equalization is growing. Furthermore, complexity and density increase is supported by evolving silicon technology able to manage more W/mm2. The realization of the transmission lines in PB even if based on the most sophisticated materials available on the market, shows very significant losses at very high operation frequency. Also connectors, even if now very optimized, contribute to increase loss and crosstalk. The electrical backplane is close to reach its limit and for that reason we need to explore alternative solutions. Due to the fact that optical technologies [2,3] are able to provide higher capacity over longer distances than electrical transmission systems, a natural answer to the limitations shown by the copper-based interconnection is to exploit optical solutions. For optical interfaces the bandwidth density scaling necessary to face the future Ethernet requirements is achieved through the reduction of the pitch size (presently to about 50μm,) and applying 2D arrays of Vertical Cavity Surface Emitting Lasers (VCSELs) and PINs [4]. Optical fibers constitute a good trade off in terms of bending losses, coupling efficiency with VCSEL sources, costs and attenuation [5]. Multimode multicore fiber and/or 2D polymer waveguides are being prototyped to match the demand [6]. The reduction of the distance at increased bit data rates for the existing multimode fibers is an important trend and continuous development is offered for example by exploiting increased modal bandwidth. Moreover, single mode low-chirp VCSELs are being developed [7] to overcome the limitations in reach due to fiber chromatic dispersion.
Last but not least novel approaches to increase bandwidth density and transmission length are being investigated, for example based on multi-level coding (PAM4, PAM16) and compact wavelength-division multiplexing systems realized in silicon photonics or InP-based photonic integrated circuit approaches. Furthermore, for short distances 850-nm solutions are supported by a huge volume of VCSELs annually produced. Single supercomputer presently absorbs millions of VCSEL-based optical links and already 60 thousands of optical links are used in a single rack. With supercomputer capacity growing (presently at 55 petaflops) further scaling in the number of optical links is expected. Exascale systems will require billions of optical links. Many studies are carrying out to better achieve the integration [8] using silica or polymeric waveguides, instead of copper striplines.
The aim of this paper is to compare the performance achievable with the best technologies available today for the electrical backplane with the ones provided by an optical backplane designed by means of fiber-based interconnections. We present a simulative analysis on different Cu-based electrical backplane architectures implemented using the technologies and materials available on the market, finding the best Cu-based backplane achievable nowadays. Moreover, we propose a fiber-based optical backplane realized exploiting VCSELs as optical sources. The most performing VCSELs proposed in literature are taken into account for our comparison. A comparative analysis in terms of capacity, power budget and consumption between the found best achievable electrical (Cu-based) backplane configuration and the designed optical solution is finally proposed.
2. Simulative analysis of performances/limits of Cu-based electrical backplane
Electrical backplane performances evolve year over year having the target to maximize both the line rate and the number of the connections. This evolution is driven by microelectronic technology and the material used for the backplane substrate. In our simulations we consider a connection from a transmitter (Tx) set on the I/O card to a receiver (Rx) on the matrix card. The I/O card is connected to the matrix board through a Cu-based electrical backplane, as shown in Fig. 3. Our target is to calculate the power consumption of this architecture putting together the transceiver data (for example the ALTERA device) and the information about the attenuation of copper link, obtained through a simulative analysis. To simulate the attenuation of the Cu-based link the commercial simulation tool HyperLynx® has been used. We have analyzed three different possible backplane configurations: the horizontal coupled striplines in the backplane, the vertical coupled striplines in the backplane and the so-called midplane. In the first geometry both the Cu coupled striplines are on the same plane. In the second one the coupled striplines are in two different plane of the backplane stackup, separated by a dielectric layer. In the midplane configuration I/O board and matrix board are orthogonally directly connected through ad hoc connectors.
A systematic series of simulations has allowed finding the best backplane configuration and I/O/matrix combination to optimize the performance of the Cu link. To achieve this goal we have considered different dielectric materials available on the market and their related roughness, the width of the Cu-based striplines and the dielectric thickness have been changed with respect to the possible maximum thickness of the backplane. We simulated different lengths of the copper striplines up to 60 cm, taking into account the presence of standards via or back drill via (the best case has been considered with a precision of ± 76 μm) in all the boards/backplane models, and Impact Molex Connectors model has been used to simulate electrical connectors. Moreover, we have taken into account that in the same layer there are only co-directional signals. The simulations have been realized considering the board temperature at 55 °C and 65 °C, according to the real operating temperatures. Table 1 resumes the best I/O board, the matrix board and the backplane characteristics used in the simulations to achieve the best condition for each configuration taken into account.
Table 1. Parameters used in simulations to achieve the best electrical configuration.
The low impedance value of the I/O board is limited to the maximum thickness of the board. Figure 4 shows the attenuation of the three different Cu-based link configurations. In terms of attenuation the midplane seems to be the best solution, but this configuration is affected by cooling problems and for this reason it has not been considered in the following power analysis. Horizontal and vertical coupled configuration curves are similar up to 15 GHz, beyond this value the horizontal coupled backplane solution is more convenient. The “holes” visible in the simulation curve are due to the residual stub in back drilled vias and connectors characteristics, which generate a discontinuity. We consider the best backplane configuration (horizontal coupled striplines) and the best match between I/O card and matrix card. For the electrical solution analysis we have taken into account the transceiver of Altera, Stratix V GT 28Gb device described in [9]. The PMA is 1.2 V supplied and the absorption depends on the operating frequency. The core chip is 0.9 V supplied. The last block of transmission buffer is 1.5 V supplied and the absorption depends only on the output swing. The reported absorption power is referred to the entire PMA block. We have considered the received power as the minimum power to guarantee error free in all the bandwidth of the receiver. In Fig. 5 we report the power analysis of the electrical solution in two cases: without and with equalization (note that the equalization function is not optimized to work at low frequencies). At low frequencies (up to 3.9 GHz) the equalization function can be switched off (the transmitted, supplied and receiver power are represented by continuous curves). To increase the useful bandwidth is necessary to activate the equalizer (blue dotted line) to compensate the distortion introduced by the Cu link, so you can reach about 11GHz bandwidth, which means approximately 20-22Gb/s transmission rate. The transmitted, supplied and received power, in case of equalization is working on, are represented by dotted lines in the picture. The results of the simulations constitute the best performance achievable nowadays with the most advanced solutions in terms of capacity and power consumption. This performance represents the reference for the solution based on optical interconnection proposed in the Section 3.
3. Analysis of a fiber-based backplane
In order to achieve a comparison with the optical solution in terms of power we have ideally designed an optical backplane by replacing (see Fig. 6) the Cu-based interconnection line with silica optical fiber (for example standard multi-mode fiber); the output buffer of electrical line driver with a VCSEL source; at last, the first input stage of receiver buffer with a photodiode.
In the fiber link we have considered fully optical connectors. The attenuation of the optical link includes insertion losses at the interfaces between Tx/Rx and optical fiber (about 0.5 dB respectively), insertion loss of 0.5 dB between ferrule-ferrule of the connector and a further loss due to the male–female connectors alignment. Hence, the total loss due to the optical link is about 4 dB. In the designed optical backplane it has been considered an optical fiber less than 1-m long, so losses due to the fiber attenuation and losses depending on the transmission frequency are negligible.
We have taken into account optical single mode and multimode VCSEL sources. VCSELs are already considered a very attractive solution for data communications. They are capable to deliver highest modulation speeds beyond 40Gb/s [10]. At the same time they consume small amounts of power and can be mass fabricated at very low cost. In order to realize energy-efficient high-speed performance, large resonance frequencies must be achieved at a low drive current [11]. Recently, large progress on energy-efficient VCSEL was made at different bit rate [12]. For our analysis we have referred to data and performances from the present literature. For example, we have considered a very efficient single-mode 850-nm VCSEL with 56-fJ dissipated energy for 25-Gb/s operation. The threshold current is 0.15 mA and the bias current is about 1 mA to achieve error-free transmission [13]. The achievable transmittable power is about −3 dBm. If we consider a commercial device to directly modulate the VCSEL, with a 3.3V supply the power consumption for the VCSEL is 5.56 dBm.
Instead, if we use a multimode 850-nm VCSEL the power consumption of the VCSEL is 6.68 dBm. The transmittable power is about 0 dBm. A V-I-Systems D30-850M photodetector can be employed with sensitivity about −9.5 dBm (BER = 10−12 at 25 Gb/s) and the received power remains always higher than sensitivity value. Hence, for very short distance typical of optical interconnects, the transmitted frequency is not limited by the backplane implementation realized by means of optical fiber, but only by the bandwidth of the exploited optical Tx/Rx devices. In our analysis we have considered the performance of only commercially available devices.
4. Comparison between the electrical and optical backplane solutions
In Table 2 we summarize the results in both solutions taken into account. In the electrical solution with no equalization we observe an upper limit of 6-7 Gb/s. With equalization it is possible to reach about 22 Gb/s, paying in terms of complexity and costs of the interfaces. It is a matter of fact that for the Cu-based electrical solution the main limitation remains the backplane line losses. In the optical solution we report the absorption power achievable in case of different VCSEL sources taken into account at different bit rate. For our comparison we use a V-I-Systems photodetector described above, with a −3-dB bandwidth of 30 GHz and we consider the consumed power to achieve error free transmission.
Table 2. Comparison between electrical and optical solution.
Regarding the transmission up to 25 Gb/s, from the calculated results for our optical design taken into account, using the data reported in Table 2 and by the comparison with the results for the electrical solution, we deduce that considering the same power budget the optical solution can easily reach 25 Gb/s only depending on the transmitter /receiver bandwidth characteristics, consuming almost a quarter of power with respect to use electrical interconnections.
As example in Table 2 we report also the performance of a multimode 850-nm VCSEL working at 40 Gb/s [14]. We obtained half power consumption per Gb/s doubling the bit rate with respect to electrical condition. Furthermore, to minimize the total power consumption in case of optical interconnection it is necessary to work on driving electronics besides the single transmitter and receiver bandwidths. Significant works continue to be done in this direction, promoting an increased chip-level of integration [15]. Anyway, in fiber-based optical solution there is no limitation due to the backplane line.
5. Conclusions
In this paper we have achieved by simulations the best performance of the optimized configuration of a Cu-based backplane exploiting the technology available today and we have realized a power analysis to compare the electrical solution with respect to a possible future optical solution based on a full optical fiber-based backplane.
In the electrical solution with equalization at a bit rate lower than about 20-22 Gb/s we achieve that the power consumption per Gb/s is higher than 0.8 mW. It is a matter of fact that for the Cu-based electrical solution the main limitation remains the backplane line losses.
Using the same power budget and considering the performance of optical Tx and Rx devices, the optical backplane can easily reach 25Gb/s only depending on Tx/Rx bandwidths. Considering the characteristics of VCSEL sources and optical receivers presented in literature, the power consumption per Gb/s is achieved lower than 0.2 mW, almost one quarter than the above explained electrical solution. There is no limitation due to the backplane line if constituted by optical fiber. In the next future, we expect significant evolution in terms of performance, power consumption and costs as soon as the optical backplane will be accepted by the industry and become widely used, accelerating the ability of the backplane to scale towards higher and higher capacity with low power consumption.
References and links
1. S. Kipp, “The Limit of Switch Bandwidth,” Proc. OFC 2011, Los Angeles, CA, paper OMV1 (2011). [CrossRef]
2. A. Taubenblatt, “Optical Interconnects for High-Performance Computing,” J. Lightwave Technol. 30(4), 448–457 (2012). [CrossRef]
3. N. Fehratovic and S. Aleksic, “Power Consumption and Scalability of Optically Switched Interconnects for High-Capacity Network Elements,” Proc. OFC 2011, Los Angeles, CA, paper JWA84 (2001).
4. D. M. Kuchta, A. V. Rylyakov, C. L. Schow, J. E. Proesel, C. Baks, C. Kocot, L. Graham, R. Johnson, G. Landry, E. Shaw, A. MacInnes, and J. Tatum, “A 55Gb/s Directly Modulated 850nm VCSEL-Based Optical Link,” Proc. IEEE Photonics Conference (IPC 2012), paper PD1.5 (2012). [CrossRef]
5. P. Boffi, A. Gatto, A. Boletti, P. Martelli, and M. Martinelli, “12.5 Gbit/s VCSEL-based transmission over legacy MMFs by centre-launching technique,” Electron. Lett. 48(20), 1289 (2012). [CrossRef]
7. N. N. Ledentsov, J. A. Lott, J.-R. Kropp, V. A. Shchukin, D. Bimberg, P. Moser, G. Fiol, A. S. Payusov, D. Molin, G. Kuyt, A. Amezcua, L. Y. Karachinskiy, S. A. Blokhin, I. I. Novikov, N. A. Maleev, C. Caspar, and R. Freund, “Progress on single mode VCSELs for data- and tele-communications,” Proc. SPIE 8276, 82760K, 82760K-11 (2012). [CrossRef]
8. C. Berger, B. J. Offrein, and M. Schmatz, “Challenges for the introduction of board-level optical interconnect technology into product development roadmaps,” Proc. SPIE 6124, 61240J, 61240J-12 (2006). [CrossRef]
9. http://www.altera.com/devices/fpga/stratix-fpgas/stratix-v/stxv-index.jsp
10. P. Westbergh, R. Safaisini, E. Haglund, B. Kögel, J. S. Gustavsson, A. Larsson, M. Geen, R. Lawrence, and A. Larsson, “High-speed 850nm VCSELs with 28GHz modulation bandwidth operating error-free up to 44Gbit/s,” Electron. Lett. 48(18), 1145–1147 (2012). [CrossRef]
11. W. Hofmann and D. Bimberg, “VCSEL-Based Light Sources—Scalability Challenges for VCSEL-Based Multi-100-Gb/s Systems,” J. Photon. 4(5), 1831–1843 (2012). [CrossRef]
12. N. N. Ledentsov, J. A. Lott, P. Wolf, P. Moser, J. R. Kropp, and D. Bimberg, “High Speed VCSELs for Energy-Efficient Data Transmission,” Proc. ISLC 2012, paper WB1 (2012). [CrossRef]
13. P. Moser, J. A. Lott, P. Wolf, G. Larisch, H. Li, N. N. Ledentsov, and D. Bimberg, “56 fJ dissipated energy per bit of oxide-confined 850 nm VCSELs operating at 25 Gbit/s,” Electron. Lett. 48(20), 1292 (2012). [CrossRef]
14. S. A. Blokhin, J. A. Lott, A. Muting, G. Fiol, N. N. Ledentsov, M. V. Maximov, A. M. Nadtochiv, V. A. Shchukin, and D. Bimberg, “Oxide-confined 850 nm VCSELs operating at bit rates up to 40 Gbit/s,” Electron. Lett. 45(10), 501 (2009). [CrossRef]
15. B. Offrein, “Silicon Photonics packaging requirements,” in Proc, IBM Silicon Photon. Workshop, Munich, Germany, 1–14 (2011). http://www.siliconphotonics.eu/munich_slides/2_IBM.pdf
