Datacentra vormen de ruggengraat van onze internetdiensten. Ze zorgen voor de opslag, de verwerking en het transport van alle informatie die wij elke dag produceren. Maar om opgewassen te zijn tegen het alsmaar toenemende dataverkeer, zullen de operatoren van de volgende generatie datacentra de snelheid van hun optische links moeten opdrijven (naar 400Gb/s) en de bandbreedte van hun datacenter-switches moeten vergroten (tot 51,2Tb/s). Imec, IDLab en de Photonics Research Group – beide imec onderzoeksgroepen aan de UGent – stellen nieuwe silicium-fotonica-technologie voor die deze evolutie kan waarmaken. De resultaten werden voorgesteld tijdens de Europese ECOC Exhibition van 2019.
Dit artikel is enkel beschikbaar in het Engels.
Optical links and switches for the next-generation datacenters
Every day, we make great use of a variety of internet services, including video streaming, web-browsing, social networking, e-mails and text messages, and cloud computing. Almost all these services run in datacenters, where a huge number of servers are connected and work together. An enormous amount of datacenter traffic travels inside and outside the datacenters. According to Cisco Global Cloud Index, the annual data center IP (or internet protocol) traffic is expected to surpass 20 Zettabytes (or 20x1021 bytes) by 2021.
Inside these datacenters, a growing number of optical links interconnect the server racks through a complex network of fiber optical cable. Today, 100Gb/s optical links – made up of 4x25Gb/s single channels or lanes – sustain the data traffic within the data center. They transmit data over several hundreds of meters of fiber, up to 2km – with single mode fiber being the optical medium of choice for spanning these large distances.
Over the next few years, datacenter operators will upgrade their networks to 400Gb/s optical links – by aggregating 4 100Gb/s lanes per link – in order to cope with the exploding demand for data communication. Because of the increasing amounts of optical links within the datacenter, these links need to be low cost and consume as little power as possible.
In parallel, the aggregate bandwidth to be handled by a single datacenter switch – which gears the information received on the input links to the appropriate output links in the datacenter fiber-optic network – will increase to 51.2Tb/s from today’s 12.8Tb/s. This calls for ultra-high-density photonics transceiver technology, tightly integrated with the switch CMOS chip. While in the past, pluggable electro-optical transceivers have been intensively used, the industry has started to embrace co-packaged optics as a possible alternative – coupling the optics and switch CMOS chip in one and the same package. This approach greatly reduces the length of the electrical lanes from the centimeter to the millimeter range, leading to improved signal integrity and speed. Co-packaged optics also promises to enable significantly higher bandwidth density and lower power consumption – meeting the scaling requirements of the next-generation datacenter interconnects.
Silicon photonics: key enabling technology
Silicon photonics (SiPho) has been widely recognized as an attractive technology for integrating the essential building blocks of optical links. With this scalable technology, compact and low-power transceivers can be implemented at low cost and at high manufacturing volume, by leveraging existing CMOS fabrication infrastructure. Passive components (including fiber couplers, waveguides and wavelength multiplexing filters) can be integrated together with active components (such as the electro-optic modulators, thermo-optic devices and Ge-based photodetectors). Electronics circuits such as drivers and trans-impedance amplifiers (TIA) can be fabricated on a separate wafer and assembled at the die level with the SiPho circuits using flip-chip assembly methods. In these circuits, the driver converts a standard CMOS bit state into an electrical voltage or current that is compatible with the optical device, while the TIA translates the photocurrent back into a standard CMOS bit state.
Imec, IDLab and the Photonics Research Group – both imec research labs at Ghent University – use imec’s Si photonics platform on 200 and 300mm SOI wafers in combination with high-speed electronics to develop the next generation SiPho interconnect technology. Recently, the teams have obtained substantial improvements at various levels of Si photonics technology development, i.e., at the level of process integration, individual component development as well as at the sub-assembly level – bringing the integrated solutions close to industrial implementation. In this article, some of these latest developments are discussed in more detail. Part of the results have been presented at the 2019 European Conference on Optical Communication (ECOC).
Towards ultra-dense Tb/s co-packaged optics: a TSV-assisted hybrid FinFET CMOS-SiPho transceiver
Realizing ultra-dense Tb/s co-packaged optics is challenging for several reasons. First, the footprint of the SiPho transmitters and receivers must be sufficiently small. Second, the SiPho optical I/O module – essentially containing the modulator, photodetector, fiber coupler, driver and TIA – needs to interconnect to the switch CMOS chip in the same package through high-density and high-speed electrical interconnects. Therefore, high-speed electrical signals should be routed to and out of the electro-optical transceiver array with low RF loss and high signal integrity. Traditionally, chip-to-package wire-bond interconnects combined with on-chip routing for RF and power/ground delivery is used. But for multi-Tb/s transceivers, this approach is unlikely to yield the required performance and density targets.
To enable high-speed electrical interconnections over short distances, the imec team has taken an alternative approach: they co-integrated 10µm wide through-Si vias (TSVs) in a 100µm thick SiPho interposer – enabling substantially improved speed and electrical interconnect density compared to the wire-bond approach. In this first-of-a-kind functional demonstrator, the TSVs route the electrical signals from the package to the 14nm FinFET CMOS die, directly through the SiPho interposer. The FinFET chip, which is flip chipped on top of the SiPho interposer, contains the modulator driver and TIA arrays, which interface directly with the optical components in the SiPho interposer. Although the TSVs were only used for low data-rate signals in this first demonstration, they have been shown to carry high frequency signals at 50GHz and beyond with low RF loss in a separate measurement.
Photographs of the TSV-assisted FinFET CMOS-SiPho transceiver prototype (as presented at 2019 ECOC).
After the hybrid assembly of the CMOS-SiPho optical module, a fiber-array was attached and optically connected to the SiPho interposer using grating couplers.
On the transmit side of the photonics transceiver, each transmitter cell – implementing a 50Gb/s ring modulator – fits in a footprint of 100µm x 100µm. The receiver cells – with Ge photodetectors and TIA – cover a similar area. At a data rate of 40Gb/s non-return to zero (NRZ), the transmitters and receivers operate with a dynamic power consumption of 0.16 and 0.08pJ/bit, respectively. In summary, the prototype demonstrated ultra-low power consumption with impressive (1Tbps/mm2) bandwidth density, paving the way to ultra-dense co-packaged optics in future datacenter switches.
Improved optical coupling: ultra-broadband low-loss single-mode fiber couplers
For applications requiring broadband optical coupling, edge couplers are preferred over grating couplers. These structures expand the optical mode diameter from sub-micron dimensions in the Si waveguides to the 10µm mode-field diameter of a single-mode optical fiber. As such, they enable efficient and broadband coupling to fibers when those are butt coupled to the edge coupler through an optical facet etched in the SiPho chip. When combined with mechanical alignment features also etched into the chip, high-throughput fiber packaging using passive alignment can be realized with the required sub-micron alignment accuracy. Such passive alignment can be beneficial over the current active alignment approaches, in which a laser needs to be coupled to the fiber array during the assembly process.
Imec recently realized low-loss edge couplers with self-aligning V-grooves. For this development a silicon nitride (SiN) waveguide layer was added to the SiPho platform. By properly etching the Si nitride layer, the optical mode can be expanded into a size that nearly matches that of the single mode fiber. A key step in imec’s design is the removal of the substrate below the buried oxide layer of the SOI wafer to prevent the expanded mode from leaking into the Si substrate. The improved design resulted in a better than -1.5dB/fiber coupling efficiency to single mode fibers for ultra-broadband operation in the O- and C-band. To facilitate passive and automated assembly of the fibers to the SiN edge couplers, self-aligned V-grooves have been developed on the chips.
(a) Schematic and (b) SEM image of SiN edge couplers (as presented at 2019 ECOC).
Enhanced optical receiver sensitivity: Ge/Si avalanche photodetectors
In order to improve the power and reach performance of optical interconnects, optical engineers are always looking for ways to lowering the optical losses across the optical path. Besides lowering optical coupling losses as described earlier, an additional option is to improve the receiver sensitivity, so the link can tolerate a higher optical loss. In order to realize such sensitivity gains, avalanche photodiodes have received considerable attention from industry over the past several years. In these diodes, an avalanche multiplication of holes or electrons – created by the initial electron-hole pairs generated by the absorbed photons – increases the gain of the diode. Such avalanche gain can help to improve sensitivity and better optical link margins, provided that the additional noise generated in the process is sufficiently low.
The imec team has developed an avalanche photodiode with a separate absorption charge multiplication (SACM) region. In this diode, Ge is used as the absorption material and Si is used as the low-noise avalanche material. In contrast to conventional SACM designs, imec’s solution has a lateral diode design with metal contacts only added to the Si region and not to the Ge region. This results in enhanced primary responsivity and reduces manufacturing complexity. The Ge/Si avalanche photodiode was fabricated using imec’s 200mm Si photonics platform. The diode features a high primary responsivity of 0.78A/W at -2V, an avalanche multiplication gain larger than 8 and bandwidth of 32GHz at -12V. Clear and wide-open eye diagrams at NRZ-OOK data rates of 25, 40 and 50Gb/s show the potential of these avalanche photodiodes in improving receiver sensitivities and optical link margins at data rates of 40Gb/s and beyond.
Electrical eye diagrams from the Ge/Si avalanche photodiode at 1550nm and -12V with NRZ-OOK data rates of 25Gb/s, 40Gb/s and 50Gb/s.
Low-power 106Gb/s PAM-4 transmitter and receiver circuits for 400GbE interconnects
The capabilities of the electro-optical transceivers – in terms of speed, power consumption and footprint – can be significantly improved by combining the Si photonics building blocks with specialized high-speed low-power electronics (including e.g. the drivers, TIAs, equalizers and converters) and higher-order signal modulation formats. This is the expertise of IDLab and the Photonics Research Group. These teams have used their specialized knowledge to efficiently co-integrate all these ingredients – which is key for obtaining industry-relevant solutions.
Using this approach, 106Gb/s low-power PAM-4 transmitters and receivers have been demonstrated. PAM-4 is the industry’s modulation format of choice for 100Gb/s per lane transmission over 500m. 400Gb/s data rates can then be obtained by using 4 of these 100Gb/s lanes. Other than NRZ, which is a 2-level modulation format, PAM-4 is a 4-level modulation format, with levels denoted as 00, 01, 10 and 11 (combining two bits in each level). While NRZ ideally switches between ‘all the light’ (1) or ‘no light’ (0), PAM-4 levels correspond to ‘no light’, ‘a third of the light’, ‘two thirds of the light’ or ‘all the light’. Hence, with this 4-level format, the data rate can be doubled while keeping the same bandwidth.
Generating and receiving PAM-4 at line rates of 100Gb/s has proven very challenging. In general, using a multi-level modulation format leads to reduced eye openings. On top of that, in a traditional single-modulator implementation, the modulator needs to have a linear transfer function. Most types of modulators, however, are non-linear devices. Attempts to solve this issue are mainly limited to the electrical domain and rely on power-hungry tools as digital signal processing (DSP) and equalization to compensate for any non-linearity.
On the transmit side: a compact and low power 106Gb/s transmitter
The team of IDLab and the Photonics Research Group realized a unique solution to enable low-power 106Gb/s PAM-4 transmission. Instead of using one single (non-linear) modulator, they propose a transmitter topology based on the addition of 2 parallel 56Gb/s NRZ electro-absorption modulators – incorporated in a Mach-Zehnder interferometer. These modulators are then wire-bonded to a 55nm SiGe BiCMOS driver.
Compared to a single-modulator solution, the transmitter now generates the PAM-4 signal optically from the two EAMs – without using any digital-to-analog converters (DACs) to go from bits to analog in the electrical domain. And this completely removes the linearity requirement from the transmitter electronics and modulators, and requires no power-hungry equalization or DSP for compensation. This innovative solution reduces complexity and results in a much lower power consumption – as low as 1.5pJ/b. 106Gb/s PAM-4 transmissions could be obtained over 1km of single mode fiber, with clear open eyes.
PAM-4 transmitter: architecture and experiment set-up (as presented at 2019 ECOC).
On the receive side: a highly sensitive and low power 106Gb/s receiver
Complementary to the transmitter, a highly sensitive and low power 106Gb/s PAM-4 receiver was demonstrated. Like the transmitter, no power-hungry digital signal processing or equalization techniques were implemented in the receiver architecture or the test bed.
Key to the solution is the use of a low-noise highly sensitive TIA. The gain and bandwidth settings of this TIA can be adjusted digitally in order to maintain good linearity for PAM-4 while preserving a high bandwidth. The TIA, implemented in 55nm SiGe BiCMOS technology, is wire-bonded to a high-speed 50GHz Si photonic Ge photodiode.
With a power consumption of only 1.51pJ/b, a compact form factor and clear open eye diagrams at 106Gb/s data rate over relevant fiber distances, this PAM-4 receiver meets all industrial requirements for 106Gb/s single-lane operation.
Imec, IDLab and the Photonics Research Group at Ghent University have tackled major technology challenges leading to Si photonics technology solutions for the next-generation datacenters. Innovations were obtained at different levels, including the process integration level (i.e., a TSV-assisted hybrid FinFET CMOS-SiPho transceiver technology demonstrator), the component level (i.e., improved edge coupler designs and a highly sensitive avalanche photodetector) and the sub-assembly level (i.e., PAM-4 transmitters and receivers). By efficiently co-integrating the individual Si photonics components with TSVs and with dedicated high-speed electronics, imec’s Si photonics technology meets major industrial requirements, bringing this mature technology close to industrial implementation.
This article was first published in Photonics Online, Nov. 2019.
Want to know more?
- Read the press release ‘Imec extends silicon photonics portfolio targeting nextgeneration datacenter interconnects’, launched at the 2019 ECOC Exhibition.
- The results presented in this article have been described in more detail in four ECOC papers, that can be requested via our contact form.