With the ongoing progress in chip scaling, the data flow to and from chip packages is increasing accordingly. The
simultaneous increase of channel count and channel speed in an essentially constant form factor becomes a more and
more demanding challenge. The resulting I/O-bottleneck is considered to be a major limiting factor for the overall
performance of future chip packages and computing systems. Optical interconnects offer both increased channel density
as well as longer link reach at high frequencies.
Our current work focuses on integrating optical I/O with standard organic packages in order to maximize the aggregate
data flow to and from such packages. We present a novel approach for attaching an electro-optical conversion module
directly on top of the organic chip package, together with experimental results of a first prototype implementation.
With the ongoing progress in chip scaling, the data flow to and from chip packages is increasing accordingly. The
simultaneous increase of channel count and channel speed in an essentially constant form factor becomes a more and
more demanding challenge. The resulting I/O-bottleneck is considered to be a major limiting factor for the overall
performance of future chip packages and computing systems. Optical interconnects offer both increased channel density
as well as longer link reach at high frequencies.
Our current work focuses on integrating optical I/O with standard organic packages in order to maximize the aggregate
data flow to and from such packages. We present a novel approach for attaching an electro-optical conversion module
directly on top of the organic chip package, together with experimental results of a first prototype implementation.
We present a novel approach for packaging high-speed opto-electronic 12x-array devices in a compact, low-cost package
for waveguide-based intra-system links. In order to avoid optical signal loss and crosstalk, the mutual alignment between
PCB-embedded multimode waveguides and the opto-electronic components needs to be in the order of 5-10 micrometer,
which is an order of magnitude tighter than standard PCB manufacturing tolerances. Our packaging concept uses a
combination of passive alignment steps, tolerance stackup reduction and a misalignment-tolerant coupling scheme in
order to bridge this gap in a cost competitive way.
Using flip-chip technology, the opto-electronic components are placed onto a very thin substrate with holes for the light
path. The top side of the 25 μm liquid crystal polymer (LCP) substrate not only provides fast and low-loss electrical
connections, but also serves as alignment reference plane for the entire module, avoiding alignment tolerance
accumulation over different assembly steps. Openings for the laser beams, passive lens alignment features, centering
holes for mechanical alignment pins between module and board and optional MT-guide receptacles are all laser-cut
within one single process step, with a precision better than 5 μm. A similar approach is used for the PCB-side optics, and
a lens-pair coupling scheme provides for a sufficiently large misalignment tolerance between the package and the PCB.
Mechanical rigidity of the package and thermal protection are provided by an epoxy filled aluminum frame.
We will present our design considerations, the basic package concept, the actual experimental implementation and
characterization results of our first prototype package.
Optical interconnects have gradually replaced electrical interconnects in the long-distance telecom, local-area, and rackto-
rack link classes. We believe that this transition will also happen in the card-backplane-card datacom link class, both for bandwidth*length reasons and for density reasons. In analogy to the transition from individually wired boards to integrated printed circuit boards, we believe that eventually board-level optical interconnects will be based on an integrated technology such as board-embedded waveguides. In order to bring optical waveguide technology into mainstream product development plans, however, numerous challenges on many levels have to be met. Problems to be tackled span from the base level of materials (stability, processability) and devices (reliability, lifetime), over the subsystem level of packages (concepts, cost-efficient assembly and alignment) all the way up to the system level (link architecture, system packaging, heat management). A sustainable solution can only be reached if the development of all individual technology components is done with the whole system in mind. Important figures of merit are the cost per gigabit per second, the power per gigabit per second, and the maturity/reliability of the technology. We will give an overview of our optical interconnect activity, with respect to these challenges. We will discuss the options, explain our technology decisions and present some results of our multi-disciplinary activity.
Polymer waveguides embedded in a printed circuit board offer a substantial increase in the achievable bandwidth density compared with today's electrical interconnects. We present our results on the polymer waveguide technology and the building blocks that perform the optoelectronic conversion. Specific challenges in integrating optics in a printed circuit board are addressed. Data transfer measurements are presented.
The development of optical interconnects in printed circuit boards (PCBs) is driven by the increasing bandwidth requirements in servers, supercomputers and switch routers. At higher data rates, electrical connections exhibit an increase in crosstalk and attenuation; which limits channel density and leads to high power dissipation. Optical interconnects may overcome these drawbacks, although open questions still need to be resolved. We have realized multimode acrylate-polymer-based waveguides on PCBs that have propagation losses below 0.04 dB/cm at a wavelength of 850 nm and 0.12 dB/cm at 980 nm. Transmission measurements at a data rate of 12.5 Gb/s over a 1-m-long waveguide show good eye openings, independent of the incoupling conditions. In the interconnect system, the transmitter and receiver arrays are flip-chip-positioned on the top of the board with turning mirrors to redirect the light. The coupling concept is based on the collimated-beam approach with microlenses in front of the waveguides and the optoelectronic components. As we aim for large two-dimensional waveguide arrays, optical crosstalk is an important parameter to be understood. Accordingly, we have measured optical crosstalk for a linear array of 12 optical channels at a pitch of 250 um. The influence of misalignment at the transmitter and the receiver side on optical crosstalk will be presented as a function of the distance between waveguide and transmitter/receiver.
Next-generation high-end data processing systems such as Internet switches or servers are approaching aggregate bandwidths in excess of 1 Terabit per second. In the case of Internet switches, the increase of fiber bandwidth that is caused by the introduction of Dense Wavelength Division Multiplexing leads to an increase of system size from single-shelf to multi-rack configurations. Intra-system interconnects will therefore span from centimeters (on-board) up to tens of meters (rack-to-rack). The task of providing hundreds of individual links at speeds in excess of 10 Gigabit per second over these distances becomes increasingly difficult for conventional copper-based technology. Using a packet switch system as an example application, we define a set of interconnect requirements for future large-scale systems. Distinguishing three interconnect classes (on-board, card-to-card over a backplane, rack-to-rack), we study the expected limits of copper-based solutions from an application point of view. After an overview of the state of the art in optical interconnect technology, we compare available technologies with the initially defined requirements. From this, we deduct key focus areas for future optical interconnect research. Finally, we present some of our recent activities in the field of waveguide and free-space based board-to-board interconnects.
We demonstrate a setup with 10 optically interconnected chips,k which can perform a distributed radix-2-butterfly calculation for fast Fourier transformation. The setup consists of a motherboard, five multi-chip-modules (MCMs, with processor/transceiver chips and laser/detector chips), four plug-on-top optics modules that provide the bi- directional optical links between the MCMs, and external control electronics. The design of the optics and optomechanics satisfies numerous real-world constraints, such as compact size (< 1 inch thick), suitability for mass-production, suitability for large arrays (up to 103 parallel channels), compatibility with standard electronics fabrication and packaging technology, and potential for active misalignment compensation by integrating MEMS technology.
We report about ongoing work on a free-space optical interconnect system, which will demonstrate a Fast Fourier Transformation calculation, distributed among six processor chips. Logically, the processors are arranged in two linear chains, where each element communicates optically with its nearest neighbors. Physically, the setup consists of a large motherboard, several multi-chip carrier modules, which hold the processor/driver chips and the optoelectronic chips (arrays of lasers and detectors), and several plug-on-top optics modules, which provide the optical links between the chip carrier modules. The system design tries to satisfy numerous constraints, such as compact size, potential for mass-production, suitability for large arrays (up to 1024 parallel channels), compatibility with standard electronics fabrication and packaging technology, potential for active misalignment compensation by integration MEMS technology, and suitability for testing different imaging topologies. We present the system architecture together with details of key components and modules, and report on first experiences with prototype modules of the setup.
We report on the optical setup, device characterization and performance in a pattern recognition task of a neural network with 256 neurons and optical feedback.
We report on ongoing work with a compact all-optical recurrent neural network with 16 X 16 channels and 256 X 256 reconfigurable interconnects (weights). We will present the optical setup and report on experimental work with the system and its building blocks. The microlens-based setup shows excellent imaging properties and easy alignability. After optimizing the setup, losses could be realized by more than an order of magnitude. The system performance is currently limited by inhomogeneities of the thresholding device.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.