Photonic Interconnects to Silicon Chips.

C. Debaesa*, M. Vervaekea, V. Baukensa, H. Ottevaerea, P. Vyncka, P. Tuteleersa, B. Volckaertsa, W. Meeusb, M. Brunfautb,

J. Van Campenhoutb, A. Hermanne a and H. Thienponta

aVrije Universiteit Brussel (VUB), Dep. of Applied Physics and Photonics (TONA/TW),

Pleinlaan 2, B-1050 Brussel bUniversiteit Gent (RUG), Dep. of Electronics and Information Systems (ELIS),

Sint Pietersnieuwstraat 41, B-9000 Gent

ABSTRACT A multi-channel free-space micro-optical module for dense MCM-level optical interconnections has been designed and fabricated. Extensive modeling proves that the module is scalable with a potential for multi-Tb/s.cm2 aggregate bit rate capacity while alignment and fabrication tolerances are compatible with present-day mass replication techniques.

The micro-optical module is an assembly of refractive lenslet-arrays and a high-quality micro-prism. Both components are prototyped using deep lithography with protons and are monolithically integrated using vacuum casting replication technique. The resulting 16-channel high optical-grade plastic module shows optical transfer efficiencies of 46% and inter-channel cross talks as low as -22 dB, sufficient to establish workable multi-channel MCM-level interconnections.

This micro-optical module was used in a feasibility demonstrator to establish intra-chip optical interconnections on a 0.6µm CMOS opto-electronic field programmable gate array. This opto-electronic chip combines fully functional digital logic, driver and receiver circuitry and flip-chipped VCSEL and detector arrays. With this test-vehicle multi-channel on-chip data-communication has been achieved for the first time to our knowledge. The bit rate per channel was limited to 10Mb/s because of the limited speed of the chip tester.

Keywords: Deep Proton Lithography, Optical interconnections, micro-optics, OE-VLSI, VCSELS.

1. INTRODUCTION For many decades the optics and photonics community has been advertising the use of optics for digital data processing. At the very beginning of this turbulent epoch the promises of embedding ultra-fast nonlinear switch [1] in parallel processing architectures created high hopes, but these devices turned out to be far too power hungry to allow massive parallelism. Around 1995 two novel paradigms for the use of optics in computing found entrance. A first paradigm features “smart pixel” arrays which consist of identical units of small CMOS processing circuitry with opto-electronic input and output ports and are primarily dedicated to local image processing or photonic switching applications [2]. The second paradigm proposes the use of optics as a wire-replacing technology at the different levels of the computer interconnect hierarchy. The strategy here is to develop or adopt advanced photonic technologies that can outperform galvanic interconnections. Meanwhile Moore’s law has been relentlessly fulfilling its prophecy, requiring a continuous growth of interconnect densities and an increase of data-throughput. Today, the conventional galvanic interconnection technology has difficulties keeping up this pace. The increased wire resistance as a result of smaller feature sizes, the residual wire capacitance, the transmission line effects and the increased inter-wire cross-talk are among the main factors that limit further advances in interconnect performances.

Since the introduction of optics as a wire replacing technology a large body of work has been dedicated to comparing the benefits and limitations of optical versus galvanic interconnections [3,4,5,6]. Typically, a “break-even length” is defined above which optical interconnects are preferred from a performance or power dissipation point of view. Although this length varies substantially with technological assumptions, a clear trend has been set: the level of system

* Christof.Debaes@vub.ac.be

hierarchy where optical interconnects show clear advantages over electrical interconnects becomes distinctly lower and is currently situated at the MCM level. By taking this trend to its extreme, there will soon be an advantage of using optics even at the off- and on-chip interconnect level. A couple of papers have already addressed the potential benefits of such an approach [3,5].

Recent advances in surface normal opto-electronic device technology and the emergence of solder-bump or related hybrid integration techniques to silicon circuits are a distinct asset to this approach and make optics an attractive candidate for the highly demanding interconnect tasks between and even on chips [7]. One of most important challenges however that remain to make such a technology practical and viable is the prototyping and low-cost manufacturing of chip-compatible, micro-Optical Interconnection Modules (OIMs) that integrate all the micro-opto-mechanical components necessary to efficiently interface these opto-electronic surface-normal transmitters and receivers.

In this paper we report on the design, the prototyping and the demonstration of such a multi-channel interconnection module. We start in section 2 by introducing the concept of this type of free-space OIMs. Next, in section 3, we provide evidence for the scalability of this approach and show that it is suitable for high density short-distance interconnects. Then, in section 4, we perform a case study by applying this design to the specific case of an opto-electronic field programmable gate array (OE-FPGA) equipped with multi-mode VCSEL and InGasAs detector arrays as photonic I/O’s. In section 5, we describe the basic principles of deep lithography with protons (DLP) applied to poly methyl methacrylate (PMMA) samples and show how we use this technique to fabricate the micro-optical building blocks of the free-space interconnect module such as the refractive micro-lens arrays, the micro-mirrors and the passive alignment features. In section 6 we briefly point out two techniques for the replication of the OIMs. We continue in section 7 to perform a simulation study of the alignment and fabrication tolerances of the component and we show that its mass-replication is possible with present-day injection molding technology. In section 8, we experimentally verify the optical throughput efficiency, the cross-talk and the alignment tolerances of a prototype OIM. In section 9, we report on the actual demonstration of intra-chip optical interconnections by combining the OIM module with the OE-FPGA and draw conclusions in section 10. The content of this invited paper is largely based on a recent journal paper [8].

2. THE CONCEPT Research groups all over the world are presently exploring various optical schemes to interconnect densely-packed photonic pin-outs regularly distributed over entire CMOS chip areas. Most of these implementations are based on beam-guiding approaches either combining or embedding rib wave-guides, individual fibers or imaging fiber-bundles into flexible [9,10] or rigid [11] modules. A second route is the use of free-space structures that use macro-optics [12,13], micro-optics [14,15] or planar optics [16] to shape and direct the light beams from transmitter to detector array. Key considerations to this approach are the interconnect distances, the relative position of the transmitters and receivers, the required interconnect densities, the power dissipation and the manufacturability of the assembly.

Recently, we have introduced micro-optical beamshaping and free-space beam-delivering structures [17] for MCM-level interconnects as shown in Figure 1. In this approach a first micro-lens array collimates the data carrying light beams emitted by the flip-chip bonded VCSEL arrays. As the beams travel through the interconnect module they are redirected towards the OE-VLSI chip by a micro-prism. Finally a second array of micro-lenses is used to focus the beams onto the detector arrays.

Total Internal Reflection

Transmitter Array Receiver Array

Micro-lens Array

Mounting Pin and holes

b)a)

Figure 1: Concept of the Optical Interconnection module (OIM). a) The OIM above a OE-VLSI Chip for intra-chip interconnections. b) A side view with a beam trace of the optical link.

Contrarily to the guided-wave approach where the cladding diameters of the individual fibers limit the maximum channel density, the free-space OIMs that we are proposing are not thwarted by this geometrical limitation but rather by the diffractive nature of light. In the next section we show that this approach is scalable and allows high density short-distance interconnections compatible with future requirements of the ITRS roadmap for MCM level and on-chip interconnections [18].

3. SCALABILITY ANALYSIS As mentioned in the previous section, one of the key considerations for free-space optical modules is the maximum obtainable channel density. This channel density strongly depends on different system parameters such as the lens diameter, the optical pathway length and the VCSEL transmitter characteristics. To understand the channel density limitations of our OIM system we started off by modeling the scaling properties of a simplified system in which the 90° deflections have been omitted. This comes down to considering a block of optical spacer material with a convex micro-lens surface at either end as shown in the schematic drawing in Figure 2. Both the detector and transmitter are separated from the OIM by a working distance d0. We characterized the VCSEL beam by a Gaussian angular intensity distribution, a beam waist 2⋅wO, a beam divergence θ and a wavelength λ0. The micro-lenses are characterized by their (back) focal length fFBL=nR/(n-1) while the spacer material between the lenses is an optically transparent plastic with index of refraction n. The distance the beam travels within this component we call the optical pathway length L. We assume the detector to be large as compared to the spot size so that only the apertures of the micro-lenses can clip the beam.

VCSEL

2⋅w 0 , θ

Lmax

2⋅w PMM A

2⋅w’0Φ lens

Detector

L

d’0d0 lens

Figure 2: The paraxial Gaussian beam propagation through a single channel system.

The first constraint that we impose on our micro-lens relay system is that of a symmetric layout. It implies that the imaged beam-waist should lie exactly in the middle of the OIM.

0' 2d L= (1)

The propagation of the VCSEL beam through this optical relay system can be described by applying the paraxial Gaussian Beam propagation method [19,20]. To that aim we have characterized the VCSEL beam with the complex beam parameter 1/q=1/R-jλ0/nπw2, where R is the beam curvature and w the beam radius. When entering the first lens

surface this beam parameter is transformed. The relationship of the beam parameter before (q0) and after the lens (q0’) is given by the following equation:

0 0 0 0 0 0 0 0 0( ) ' ( ' ) ' ( ) 'f nd q f d nq q q n f nd d nd f− − − = + − +

At the beam-waist (R=∞) the complex parameter q is purely imaginary, leaving only imaginary terms in the left hand side of the equation and gathering all the real terms in the right hand side. The transformation of a paraxial Gaussian beam can thus be described by the following two real equations:

( ) 0'''2

000

2000 =

−+− wwnfdddfn

λπ (2)

( ) ffdwwnd +−= 02

0

20

0 ''

(3)

To calculate the optical transmission efficiency of this system, we calculate the power transmission through two consecutive apertures, which is done by integrating the radial intensity over the aperture surfaces. An aperture with a diameter that is three times the beam waist will pass 99% of the optical power. Hence, to obtain a system with an overall system transmission efficiency of 98%, we require that the lens diameter should be at least 3 times the beam radius. This imposes the following constraint:

2

20

000 133

+==Φ

wdwwlenslens π

λ (4)

This lens diameter Φlens will straightforwardly determine the channel density of our OIM. The set of four equations (1)-(4) now describes the paraxial Gaussian beam propagation of a single channel system. We have four non-linear equations to solve five unknowns (f, Φlens, d0, d’0, w’0). This leaves us with an under-constrained problem and allows us to pursue several options in trying to optimize the design.

In a first attempt we propose that L should be twice the Rayleigh distance zR= nπw’02

/λ0 =L/2. This fixes the value of the secondary beam waist 2⋅w’0 and will result in a design yielding a minimal lens-diameter [20]. The beam waist will increase with a factor 2 over the Rayleigh distance L/2 so that the lens diameter will become:

' 003 3 2 3lens lensw w L

nλπ

Φ = = = ⋅

In this model, which we call the Rayleigh approach, the lens diameter Φlens is independent of the divergence angle of the input laser. However, by carrying out a more elaborate radiometric ray-trace simulation using a Monte-Carlo scheme where rays are emitted from the source in a quasi-random manner in accordance with the Gaussian emission probabilities, we found that the latter is not completely true. Indeed, highly divergent VCSELs will need micro-lenses with smaller focal lengths. As a consequence the optical system will start suffering from larger spherical aberrations. As can be seen in Figure 3, the simulated optical efficiency is dropping dramatically below the theoretical 98% for VCSELs that have FWHM divergences above 7º. Note that for such a system, an increase in lens diameter, while keeping all the other parameters identical, would only marginally increase the efficiency, since most of the aberrated light will still not be confined within the apertures of the system. We can conclude that although for every given interconnection length this first approach would result in a theoretically minimal lens diameter and thus a maximal channel density, it suffers from a significant drop in transmission efficiency for VCSELS with FWHM divergences above 7º.

(L = 8 mm)

2 0

3 0

4 0

50

6 0

70

8 0

9 0

10 0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

FWHM VCSEL (°)

Opt

ical

effi

cien

cy

(%)

D iamet er 12 3 umD iamet er 2 0 0 umPred ict ed Ef f . ( 9 8 %)

Figure 3: Optical efficiency in a Rayleigh system with micro-lenses of 123 µm and 200 µm diameter as a function of the divergence angle of the VCSEL for an interconnection pathway length of 8 mm

0

500

1000

1500

2000

1 2 3 4 5 6 7 8 9 1011121314151617181920

Pathway length (mm)

Lens

dia

met

er (µ

m) FWHM 12° 8° 6°

Rayleigh System

Figure 4: The required lens diameters as a function of the pathway length for different VCSEL divergences in a 4-f system configuration and in a Rayleigh configuration.

In a second attempt to optimize the VCSEL-detector link we studied the scaling properties of a 4-f system. Here, the working distance d0 equals the front focal length of the micro-lens (d0=fFFL=fBFL/n), and the pathway length L equals twice its back focal length (d0’=L/2=fBFL). Substituting this into Eq. (4) results in:

2

20

00 2

13

+=Φ

nwLwlens π

λ

In this 4f-system, in contrast with the “Rayleigh approach” the lens diameter strongly depends on the VCSEL divergence angle (since θ ÷ 1/w0) for a fixed working distance. The reason for this is that since the focal length is fixed for a given pathway length, we need larger micro-lens diameters for larger divergent VCSELs. This approach therefore requires significantly larger lens-diameters as can be seen in Figure 4, resulting in unacceptably low channel densities.

For moderately to highly divergent VCSELs (θFWHM>7°) the two previous approaches resulted in designs that are not in agreement with the systems performance requirements. Therefore we have investigated a more general system where neither the lens diameter nor the focal length are directly linked to the path length as was the case in the former two approaches. In this approach we either take the focal length or the diameter of the lens as a free parameter and solve the remaining parameters with the set of equations 1 to 4.

After calculating the geometrical and optical characteristics of these symmetric optical data link systems for VCSELs with FWHM divergences ranging from 1° to 20° and for path lengths between 4 mm and 20 mm, we have determined their optical collection efficiency via radiometric simulations. This allowed us to find the minimal lens diameter for a specified optical throughput efficiency depending on the pathway length and on the VCSEL divergence. The results are summarized in Figure 5 for a “low-divergence” system based on VCSELs with FWHM divergences up to 7° where spherical aberration has no significant impact (low divergence VCSEL system) and for a system equipped with VCSELs featuring a divergence of 12° where spherical aberration does become important (high divergence VCSEL system). The postulated efficiency threshold is 85% while the ratio between the channel pitch and the lens diameter was chosen such that the cross-talk between neighbouring channels was kept below –25 dB.

The use of micro-lenses with a diameter of 100 µm (point A in Figure 5) for example allows to bridge a pathway length Lmax of 5.3 mm both for VCSELs with a FWHM smaller than 7° and for VCSELs with a FWHM of 12°. This size of micro-lenses results in a channel density of 104/cm2 for the low-divergence system and only 4740/cm2 for the system with highly divergent VCSELs (point B) because of cross-talk considerations. When we use micro-lenses with a diameter of 200 µm we can bridge a length of 21.1mm (point C) in case of the low-divergence VCSELs as compared to

only a length of 13.2mm (point D) with the high-divergence VCSELs. The channel densities corresponding with this 200 µm lens diameter are 2500/cm2 (point E) and 1192/cm2 (point F) for the low-divergence VCSEL and the high-divergence VCSELs respectively.

Figure 5 therefore summarizes the scaling behaviour of our spherical micro-lens based interconnection module as a function of the maximum pathway length considering the aberrations in the system. Once the lens-diameter is chosen, all other system parameters are fixed.

Considering the high bandwidth, low latency performances of present-day VCSELs and the type of interconnect densities that we can achieve with this approach for pathway lengths below 20 mm we are definitely matching future requirements of the ITRS roadmap [18] for intra-chip interconnections.

To further extend the throw we can bridge with this kind of free-space OIM to MCM-level distances, we can e.g. compensate for spherical aberration by optimizing the micro-lens shape. This will result in higher optical efficiencies without increasing the lens diameter and will lead to a better scaling behaviour of the relay system than predicted by the paraxial Gaussian beam propagation. Also in this case we have considered different design layouts and we have optimized the micro-lens shape for systems with path lengths between 4 mm and 20 mm using VCSELs with FWHM divergences ranging from 1° to 20°. The optimization of the micro-lens shape resulted in hyperbolic lenses. We have evaluated these designs both via radiometric and beam propagation methods and we found that the optical efficiency of an optimized system is higher than 95% even for VCSELs with divergences θFWHM up to 20°. This means that when using optimized hyperbolic micro-lenses the ideal scaling behaviour of the Rayleigh system can be achieved. In Figure 6 we compare the scaling behavior of micro-optical relay systems with both spherical and hyperbolic lenses for VCSELs with θFWHM = 12°. From this graph we can conclude that implementing hyperbolic micro-lenses results in the minimum lens diameter and therefore in an optimum channel density.

1

10

100

1000

1 10 100Maximum path length Lmax (mm)

Lens

dia

met

er (µ

m)

A

B

C

105

104

103

102

Channel density (1/cm

²)

D

FWHM < 7°

FWHM < 7°

FWHM = 12°

FWHM = 12°

E

F

Figure 5: Comparison of the scaling laws of a system based on low divergence VCSELs that is not affect by aberration and a system where the VCSEL has a FWHM divergence of 12°.

1

10

100

1000

1 10 100Max Pathway length (mm)

Lens

Dia

met

er (u

m)

Channel density (1/cm

²)

105

104

103

102

hyperbolic

spherical

spherical

hyperbolic

Figure 6: Comparison of the scaling behaviour of the OIM with spherical micro-lenses and hyperbolic micro-lenses for a VCSEL with qFWHM = 12°.

By using hyperbolically shaped micro-lenses we can thus extend the range of the free-space high channel-density interconnections (≥103 cm-2) to pathway lengths well into the centimeter regime. This, combined with the high speed characteristics of present-day VCSELs (>1Gb/s), makes that the OIM is scalable well into the multi-Tb/s⋅cm2 aggregate bit rate capacity regime, which is a good match for the future requirements of the ITRS roadmap [18].

4. DESIGN OF THE OIM In the previous section we modeled the scaling behavior of a free-space multi-channel micro-optical relay module for both spherical and hyperbolic micro-lenses. We showed that both the interconnection length and the beam divergence determine the minimum required micro-lens diameter, which in its turn sets a limit to the channel density and the aggregate throughput. In this section we apply the design with spherical micro-lenses to a specific example, namely that

of an OE-FPGA intra-chip interconnection link. The OE-FPGA chip we used was designed within the framework of the project Optically Interconnect Integrated Circuits (OIIC) where a consortium of partners has been working towards a manufacturable solution for optical interconnects between CMOS IC's [21]. The dedicated FPGA CMOS chip [22,23] is equipped with opto-electronic components that are flip-chip mounted as shown in Figure 7. As emitters Multimode (MM) through-substrate emitting VCSELs were chosen with an average threshold current of 0.8 mA at 1.45 V. These devices exhibit peak conversion efficiencies of 37% at an optical output power of 3 mW at 980 nm. They have an aperture diameter of 7 µm and their far-field intensity pattern has a FWHM divergence angle of 12º, corresponding to a 1/e2 divergence angle of 10.2º. This angle is about two times larger than what is predicted when we assume that the laser would produce a single mode diffraction-limited beam. This suggests that the laser is oscillating primarily in the fundamental mode (TM00) and the two first order modes (TM01 and TM10).

By using elaborate beam-propagation methods (BPM) to describe the propagation of the above modes, we have found that the VCSEL can equivalently be modeled by a Gaussian emittance with a beam-waist of about two times the diameter aperture for an equivalent far-field pattern. To further verify our source model, we simulated the complete OIM system with the BPM method, which is propagating all the modes individually, and obtained the same results as with the radiometric simulations within a margin of 4%.

The hybrid InGaAs/InP detectors have responsivities of 0.69 A/W with variations smaller than 2% over the entire array. They have a circular aperture of 150 µm and are recognizable in Figure 7 by the cylindrical substrate removal etches. Both the VCSELs and the detectors feature a 250 µm device pitch.

To optimize the design of the module with geometrical dimensions as shown in Figure 8, we performed a set of radiometric simulations on the module for lenses with different focal lengths f and working distances d0. We determined the optical throughput efficiencies by calculating the ratio of the number of rays that fall on the detector area and the number of rays emitted by the VCSEL source. The resulting plot for various working distances and focal lengths is depicted in Figure 9. We have found that the optimal design would allow an 83% optical transmission efficiency for an optical module with spherical lenses that have a 520 µm focal length and a slightly lower working distance of 510 µm.

Figure 7: Photograph of the wire-bonded OE-FPGA chip with four optical arrays flip-chip mounted on the CMOS chip.

1mm 1mm 2.75mm

1mm

5.5mm 1mm

0.75mm

500µm

1.25mm

250µm (pitch)

200µm (diameter)

750µm

45°

Focal length: 520µm Working distance: 510µmPathway length: 8mm

Figure 8 : Dimensions of the microoptical interconnection module.

The remaining 17% of the light is not falling on the detector, suggesting that the cross-talk between neighboring channels could be significant. For the above design with a 250 µm pitch and 200 µm diameter micro-lenses however, we found the cross-talk contribution of each channel to its nearest neighbour to be -40dB which is sufficient for a practical receiver design.

40

50

60

70

80

90

100

400 450 500 550 600 650 700Working distance

Opt

ical

effi

cien

cy (%

)

Envelope

f=480µm,

Working Distance (µm)

520µm, 540µm

580µm 620µm

Figure 9: Optical collection efficiency of the OIM for various focal lengths and working distances.

5. FABRICATION WITH DEEP PROTON LITHOGRAPHY Deep Lithography with Protons [24,25] (DLP) is a dedicated technology for the rapid prototyping of three-dimensional micro-opto-mechanical systems. With this technology, different optical structures and components can be realized in one block to form monolithic micro-optical systems. Its concept is somewhat similar to that of LIGA [26] but it is based on the use of ions rather than electromagnetic radiation to structure and shape the PMMA-material. The fabrication process consists of three basic procedures: the patterning of polymethyl methacrylate (PMMA) samples with proton irradiation followed by either an etch removal of the irradiated regions with a specific developer or a swelling procedure involving a diffusion of an organic monomer vapor (see Figure 10). If required both processes can be applied to the same sample.

Figure 10: Basic fabrication processes of deep lithography with protons. After a patterned irradiation we can either apply binary chemical solvent to etch away irradiated regions or in-diffuse a monomer vapor to create micro-lenses.

PMMA wafer

Mask

Proton beam

MMA Vapor

Solvent

A) Irradiation with high-energy ions The principle of the DLP process is based on the fact that ions transfer energy to the PMMA molecules while propagating in the substrate. These interactions cause molecular chain scissions, reduce the molecular weight of the polymer and change the chemical properties of the material. In our experiments we have used protons with a specific energy of 8.3 MeV making it possible to cut through 500 µm thick PMMA samples. By accurately controlling the dose of the irradiated zones we can engineer the change in molecular weight such that we make the irradiated zones susceptible to either a binary chemical solvent or an in-diffusion of a monomer vapor.

B). Selective Etching: Micro-drilling optical quality surfaces The irradiated domains can be selectively dissolved with a special solvent, as they show higher solubility than the non-irradiated domains. In this way complex structures can be micro-drilled or precision-cut with high optical quality. The flatness and the roughness of the resulting surfaces are determined by the magnitude of the straggling effect of the impinging protons, by the precision of the movement of the translation stages and by the homogeneity of the deposited dose. In recent years we have improved both our irradiation and etching process such that at present we can fabricate high-quality surfaces with an optical flatness of λ/10 over a length of 2.5 mm and an RMS roughness of 20 nm.

With this technique we have prototyped micro-prisms with dimensions as described in Figure 11. The surfaces of both micro-mirrors constituting the top-prism of the OIM have been characterized with a vertical scanning non-contact optical profiler resulting in a surface roughness of about 32.4 nm and the flatness along the length of the micro-mirrors is better than 1 µm over a distance of 300 µm.

a) b)

c)

Figure 11: The top prism with its mounting features (a) before and (b) detached form the substrate after the etching process (c) Base plate with alignment holes and integrated micro-lenses

C). Monomer in-diffusion: 2D-spherical micro-lenses As an alternative to dissolving the irradiated zones we can swell them by exposing them to an organic MMA monomer vapor. Indeed, when regions feature a low enough molecular weight they can be receptive to an in-diffusion process of an organic monomer upon which their volume will expand. This way an irradiated regions with a circular footprint will be transformed into hemi-spherically shaped micro-lenses [27]. A thermal polymerization procedure finally prevents the out-diffusion of the monomers and fixes the shape of the micro-lenses. The physics and the details on the technological processing steps behind this technique have been published elsewhere [17,24].

Using the above technique we fabricated a base plate which contains two 2x8 arrays of 200µm diameter micro-lenses. However their measured focal lengths turned out to be 790 µm rather then the targeted 510 µm, implying that the component will exhibit lower transmission efficiencies than predicted by the optimal design. The micro-prism and the micro-lens arrays were assembled by plugging the prism into the base plate using two alignment holes. This way we obtained the prototype OIM.

6. TOWARDS LOW-COST MASS-FABRICATION: REPLICATION TECHNIQUES. Although DLP is an attractive rapid prototyping technique for micro-optical components it is not well suited for low-cost mass-production. It is however possible to use the prototype as a master and replicate the latter with a number of techniques. The most suitable candidate for replicating elements in huge quantities is injection molding whereas vacuum casting techniques can be used when copies are needed in much smaller quantities. For the injection molding process a mold from the master component can be fabricated by electroplating it using a galvanoforming process similar to that used in the LIGA-technique [26]. With this process an inverse of the master is obtained which subsequently can be used as a mold for mass-replication in a wide range of optical-grade plastics [28].

Figure 12: A side-view photograph of the replica of the OIM with its mounting structure. The two 2x8 arrays of micro-lenses are clearly visible beneath the micro-prism

For the OE-FPGA on-chip interconnect demonstrator we replicated the assembled optical interconnection module and its mounting structure via a vacuum casting technique. Here, a rubber mold is made starting from the original component and is then used to produce a limited number of copies in a high optical quality polyurethane. The resulting OIM is a monolithic bloc integrating the micro-lens arrays and the prism (see Figure 12).

7. ALIGNMENT AND FABRICATION TOLERANCES To assess the manufacturability of the design we derived the tolerance requirements for the molded component through a sensitivity analysis based on the ray-tracing models and the radiometric calculations we mentioned in sections 3 and 4. Our sensitivity analysis is based on the calculation of the tolerance to manufacturing errors and mechanical misalignments in terms of the degradation of the optical throughput and the increase of inter-channel cross-talk. To determine the tolerance for each parameter individually we have defined a minimum system performance threshold: the cross-talk should stay below –25 dB and the efficiency drop should not be larger than 15%. Thus, the tolerance on a parameter is either limited by a too large drop in throughput efficiency or by an unacceptable rise in cross-talk. We categorized the misalignments in two groups as shown in Table 1. A first group summarizes the fabrication errors of the OIM assembly and a second group shows the possible positional misalignments of the OIM above the OE-VLSI chip. In Figure 13 a schematic view of two rotational fabrication errors are given; a) an error in the top-angle (dα) of the micro-prism and, b) an unwanted tilt (dβ) of the micro-prism. The graphs illustrate both the drop in efficiency and the increase in cross-talk for these rotational fabrication errors, using a multi-mode VCSEL source as described in section 4. Other fabrication tolerances that were studied are: dYp,b a lateral shift of the prism with respect to the base plate, df a change in the focal length of the lenses and dZoim a change in the optical pathway length in the OIM. The positional misalignment errors depend on the relative position of the OIM with the OE-FPGA chip. We studied rotational misalignments (dXrot and dYrot), positional misalignments (dXPos and dYpos) and the sensitivity dependence to a deviation from the nominal working distance (dZpos). The results of the complete sensitivity analysis are summarized in Table 1, where the perturbations are ranked in order of importance from 1 to 8, such that the must stringent tolerance parameter is numbered 1.

5055606570758085

-60 -45 -30 -15 0 15 30 45 60Tilt of prism (Arc min)

Effic

ienc

y (%

)

-5-10 -15 -20 -25 -30

X-talk (dB)

5055606570758085

-45 -30 -15 0 15 30 45Deviation on top angle (Arc min)

Effic

ienc

y (%

)

-5-10 -15 -20 -25 -30

X-talk (dB)

Deviation on top angle (arc min)

dα dβ

a) b)

Figure 13: Sensitivity analysis of the rotational misalignments of the optical throughput and cross-talk

depending on a) the top-angle of the prism with MM-VCSEL b) the tilt of prism of MM-VCSEL

The sensitivity analysis was carried out for the optical relay system with both multi-mode (MM) VCSELs (θFWHM = 12°), which were flip-chipped on the OE-FPGA chip, and single-mode (SM) VCSELs with the same active area (θFWHM = 7.3°). For comparison, the simulation for the single mode source was performed with both a radiometric ray-trace (RAD) method as well as a beam propagation method (BPM). A more in-depth study on the simulation result of the alignment tolerances has been published elsewhere [8]. The extensive modeling on all alignment parameters allowed us to conclude that 1) the component can be fabricated in large copies with present-day injection molding techniques, 2) the position of OIM above the opto-electronic chip is critical, positional errors should stay below 4 to 7 µm and rotational errors should be below half a degree, 3) emitters with low divergence and single mode profiles allow for OIMs with superior tolerance values as opposed to high divergence, multi-mode emitters.

Tolerance parameter SM (RAD)

SM (BPM)

MM (RAD)

Achievable with injection

mold tech.

Ran

k

Fabrication errors

Lateral misalignm. Position prism/baseplate dYpb ± 32 µm ± 29 µm ± 19 µm ± 10 µm 3

Topangle prism dα ± 0.45° ± 0.42° ± 0.27° ± 0.083° 5 Rotational misalignm. Tilt of prism dβ ± 0.83° ± 0.75° ± 0.5° ± 0.083° 6

Focal length lens df ± 40 µm ± 40 µm ± 35 µm ± 25 µm 7 Longitudinal misalignm. Path length in bridge dZoim ± 500 µm ± 200 µm 8

Positioning errors Around X axis dXrot ± 0.85° ± 0.80° ± 0.30° Rotational

misalignm. Around Y axis dYrot ± 0.70° ± 0.65° ± 0.30° 1

Pos. em./det. Array dXed ± 7 µm ± 6 µm ± 4 µm Lateral misalignm. Pos. em./det. Array dYed ± 7 µm ± 6 µm ± 4 µm 2

Longitudional misalignm. Working distance dZa ± 40 µm ± 40 µm ± 40 µm 4

Table 1: Comparison of the required tolerances for SM- and MM-VCSELs as a result of the RAD and the BPM simulations.

8. EXPERIMENTAL VERIFICATION The master OIM was also optically characterized with an experimental set-up where the emission characteristics of the VCSEL source were mimicked with a single mode input fiber (NA=0.11) connected to an 850 nm semiconductor laser. A multimode fiber with a NA=0.2 and a core diameter of 50 µm connected to an optical power meter simulates the photo-detector. Two automated translation stages and a specially developed alignment algorithm are used to scan and seek for the optimal position of the fibers for both the input and the output channel as schematically shown in Figure 14.

The optical transfer efficiency for the 16 channels was measured to be in the range of 40% to 46% with a cross-talk between neighboring channels of –22 dB to –27 dB. Although the experimentally obtained optical transfer efficiencies are sufficiently high and the cross-talk sufficiently low to obtain reliable short distance parallel data transfer, it is clear this first prototype does not reach the calculated transmission efficiency of 83% and the –40dB cross-talk. We can impute this partly to geometrical and optical fabrication imperfections, partly to misalignment errors and partly to the fact that the input and output fibers are only approximations of the VCSEL and detector characteristics. With this experimental set-up we were also able to verify some of the assembly tolerances by moving the input fiber away from its nominal position in the lateral and longitudinal directions. Therefore we have first aligned the fibers to a position with maximal transfer efficiency, and subsequently measured the optical power for different misalignments. The cross-talk was obtained in a similar way by measuring the optical power coupled into adjacent channels. As an example, Figure 15 shows both the experimental and the modeling results for lateral misalignments.

The experimental results show a tighter tolerance than the ones derived by optical simulation. We found a ±2 µm lateral tolerance, limited by the drop in throughput efficiency. The longitudinal misalignment tolerance as predicted by simulation is much more relaxed. We measured a tolerance range of ±40µm limited by cross-talk.

Input Micro-lens

Output Micro-lens

Input fibre circular scan f1 freq.

Output fibre circular scan f2 freq.X

Y Z

Figure 14: Experimental verification system. A single mode fiber (NA=0.11) is mimicking the

input and a multi mode fiber (NA=0.2) is simulating the chip detector.

0102030405060708090

100

-20 -15 -10 -5 0 5 10 15 20Lateral misalignment (µm)

Effic

ienc

y (%

)

exp. eff. (%) exp. Xtalk (dB) simul eff (%) simul Xtalk (dB)

Cross talk (dB

)

0-10 -20 -30 -40 -50

Figure 15: Comparison of simulation and experimental results for the optical efficiency and cross-talk as a function of the

lateral misalignment.

9. MULTI-CHANNEL ON-CHIP INTERCONNECT DEMONSTRATION OVER THE OE-FPGA In this section we report on actual link experiments of the OIM above the OE-FPGA. We opted for an OE-FPGA as test vehicle because of its relevance in the introduction of area optical interconnections for inter- and intra-MCM data exchange [17].

An FPGA usually consists of a regular array of configurable logic blocks with programmable combinatorial functions while the array of logic blocks is intertwined with a configurable network of routing channels. This flexibility of the FPGA comes however at a price: the size of designs an FPGA can harbor are much lower than ASIC implementations and FPGAs are generally plagued with limited programmable routing resources. This lack of interconnect capacity, either intra-FPGA and inter-FPGA, is a well-known bottleneck for such applications.

Within the EC "OIIC" consortium a full-custom FPGA [29,30] has been developed with area optical input/output ports. In the 0.6µm CMOS design, the switching nodes can be configured to route two signals through an optical input output port. Therefore, each switching node consists of a quadruplet of two optical transmitters and two receivers. The 8×8 receiver circuits developed for the system demonstrator consist of AC-coupled low-noise trans-impedance stages and an analog/digital converters (limiting amplifier). The dynamic range of the receivers is 10dB for a consumption of 15mW at the nominal line rate of 160Mbit/s. The integrated CMOS driver and receiver circuits are interspersed with digital hardware. Manchester encoded links are used that allow a 160 Mbaud signaling rate, providing an 80 Mbits/s bit data rate.

Combining an optical pathway block with the OE-FPGA as shown in Figure 16, we demonstrated for the first time a multi-channel intra-chip optical interconnection. Figure 17 gives an overview of the results obtained with 4 adjacent channels. The lowest trace toggles at 10 Mbit/s, the above trace oscillates at half that speed while the upper traces are obtained with a logic function implemented with the FPGA such that the produced “001”-sequences are shifted. The signals were transmitted through the OIM with no apparent cross-talk The Tektronix LV500 chip tester limited the speed to 10 Mbit/s. It is possible however to increase the speed with a factor of 10 by reprogramming the board on the tester at the expense of the flexibility of the measurement setup.

Figure 16: The micro-optical interconnection module aligned with the OE-FPGA.

Figure 17: Four channel intra-chip optical data-transmission. The lower signal channels were

programmed to toggle at respectively the clock frequency (10Mhz) and half the clock frequency. The upper channels were programmed to produce shifted

“001”- sequences.

10. CONCLUSIONS

In this paper we reported on the concept of a scalable multi-channel OIM with the potential of multiple Tb/s.cm2 aggregate bit rate capacity for on-chip and MCM-level interconnection distances. A prototype module was designed to work with an OE-FPGA. The individual components were fabricated with deep proton lithography in PMMA and their optical characteristics were measured. Although experimental optical throughput efficiencies of 40-46% and cross-talk of −25dB have some room for improvement as predicted by our simulations, thus performance proved to be sufficient to make a reliable optical interconnection link demonstrator.

To assess the manufacturability of the assembly we simulated the most important fabrication and alignment tolerances, which were found to be within the reach of today’s commercial available mass-fabrication techniques such as injection molding.

Finally, we demonstrated for the first time to our knowledge a proof-of-principle intra-chip multi-channel optical interconnect on an OE-FPGA. The structure was set up to link 4 adjacent channels at 10 Mbit/s per channel.

ACKNOWLEDGEMENTS This work was funded by the European Commission ESPRIT-MELARI project 22641 ‘OIIC’, by DWTC IAP Photon network, by the FWO, IWT-GBOU, GOA and the OZR of the Vrije Universiteit Brussel. C. Debaes is indebted to the FWO for his post-doctoral fellowship.

REFERENCES [1] B. S. Wherret, “Semiconductor optical bistability: Towards the optical computer”, Springer Proceedings in

Physics 7, pp. 180-221, Springer-vVrlag, Berlin, 1985. [2] H. Scott Hinton, An introduction to Photonic Switching Fabrics, New York and London, Plenum Press,1993 [3] D.A.B. Miller, “Rationale and Challenges for Optical Interconnects to Electronic Chips”, Proceeding of the IEEE,

Vol. 88, No. 6 (June 2000), pp. 728-749. [4] M. R. Feldman, S. C. Esener, C. C. Guest, and S. H. Lee, “Comparison between optical and electrical

interconnections based on power and speed considerations," in Applied Optics, Vol. 27 (1988), pp. 1742-1751. [5] G. I. Yayla, P. J. Marchand, and S. C. Esener, “Speed and energy analysis of digital interconnections: comarison

of on-chip, off -chip, and free-space technologies,” in Applied Optics, Vol. 37, No. 2 (1998), pp. 205-227. [6] A. Krishnamoorthy and D. Miller, “Scaling Opto-electronic-VLSI Circuits into the 21st Century: A Tech- nology

Roadmap,” in IEEE J. Selected Topics in Quantum Electronics, Vol. 2 (Apr. 1996), pp. 55-76. [7] A. H. Guenther, “International Trends in Applied optics”, Bellingham, WA, SPIE Press, 2002, ch. 7. [8] C. Debaes, M. Vervaeke, V. Baukens, H. Ottevaere, P. Vynck, P. Tuteleers, B. Volckaerts, W. Meeus, M.

Brunfaut, J. Van Campenhout, A. Hermanne and H. Thienpont, “Low-cost micro-optical modules for MCM level optical interconnects”, IEEE J. Selected Topics in Quantum Electronics, Vol. 9, No. 2 (2003), pp. 518

[9] A. Neyer, B. Wittmann M. Jöhnck, “Plastic-optical-fiber-based parallel optical interconnects”, IEEE J. Selected Topics in Quantum Electronics, Vol. 5, No 2 (1995), pp 193-200.

[10] R. Bockstaele, A. Van Hove, T. Coosemans, C. Sys, I. Moerman, B. Dhoedt, R. Baets, P. Van Daele, J. Koetsem and L. Van der Torren, “Microcavity LED-based parallel data link using small-diameter (125 µm) plastic optical fibres”, J. Opt. A: Pure Appl. Opt., Vol. 1, No. 2 (1999), pp. 233-236.

[11] Y. Li, J. Popelek, L. J. Wang, Y. Takiguchi, T. Wang, K. Shum, “Clock delivery using laminated polymer fiber circuits”, J. Optics A: Pure and Applied Optics, Vol. 1, No. 2 (1999), pp. 239-243.

[12] M.W. Haney, M.P. Christensen, P. Milojkovic, G.J. Fokken, M. Vickberg, B.K. Gilbert, J. Rieve, J. Ekman, P. Chandramani, F. Kiamilev, “Description and evaluation of the FAST-Net smart pixel-based optical interconnection prototype”, Proceedings of the IEEE, Special issue on Optical interconnections for digital systems, Guest Ed. Y. Li, E. Towe, M.W. Haney, Vol. 88, No. 6 (2000), pp. 819-828.

[13] M.W. Haney, M.P. Christensen, P. Milojkovic, G.J. Fokken, M. Vickberg, B.K. Gilbert, J. Rieve, J. Ekman, P. Chandramani, F. Kiamilev, “Description and evaluation of the FAST-Net smart pixel-based optical interconnection prototype”, Proceedings of the IEEE, Special issue on Optical interconnections for digital systems, Guest Ed. Y. Li, E. Towe, M.W. Haney, Vol. 88, No. 6 (2000), pp. 819-828.

[14] D. T. Neilson, E. Schenfeld, “Plastic modules for free-space optical interconnects”, Applied Optics, Vol. 37, No. 14 (1998), pp. 2944-2952.

[15] G. Kim, X. Han, R.T. Chen, “Crosstalk and interconnection distance considerations for board-to-board optical interconnects using 2-D VCSEL and microlens array”, IEEE Photonics Technology Leters., Vol. 12, No. 6 (2000), pp. 743-745.

[16] S. Sinzinger, J. Jahns, “Integrated Micro-Optical Imaging-System With A High Interconnection Capacity Fabricated In Planar Optics”, Applied Optics, Vol. 36, (1997) pp. 4729-4735.

[17] H. Thienpont, C. Debaes, V. Baukens, H. Ottevaere, P. Vynck, P. Tuteleleers, G. Verschaffelt, B. Volckaerts, A. Hermanne and M. Hanney, “Plastic Microoptical Interconnection Modules for Parallel Free-space Inter- and Intra-MCM Data Communication”, Proceedings of the IEEE, Vol. 88, No. 6 (June 2000), pp. 769-779.

[18] International Technology Roadmap for Semiconductors, 2001 edition, [online] Available: http://public.itrs.net/Files/2001ITRS/Home.htm

[19] H. Kogelnik and T. Li, “Laser Beams and Resonsators”, Proceedings of the IEEE, vol 54, pp. 1312-1329, 1966 [20] E. Siegman, “Complex paraxial wave optics”, Chapter 20, in Lasers, University Science Books, 1986, Mill Valley,

CA

[21] Optically Interconnected Integrated Circuits, Rationale and Project summary, [online]. Available: http://www.intec.rug.ac.be/oiic/rationale.html

[22] M. Brunfaut, J. Depreitere, W. Meeus, J. Van Campenhout, H. Melchior, R. Annen, P. Zenklusen, R. Bockstaele, L. Vanwassenhove, J. Hall, A. Neyer, B. Wittmann, P. Heremans, J. Van Koetsem, R. King, H. Thienpont, and R. Baets, “A multi-FPGA demonstrator with POF-based optical area interconnect," in Proceeding of the IEEE Laser and Electro-optics Society conference, Vol. 2, pp. 625-626, 8 Nov. 1999.

[23] M. Brunfaut, W. Meeus, J. Van Campenhout, R. Annen, P. Zenklusen, H. Melchior, R. Bockstaele, L. Vanwassenhove, J. Hall, B. Wittman, A. Nayer, P. Heremans, J. Van Koetsem, R. King, H. Thienpont, and R. Baets, “Demonstrating optoelectronic interconnect in a FPGA based prototype system using ip chip mounted 2D arrays of optical components and 2D POF-ribbon arrays as optical pathways," in Proceedings of the SPIE, 4455, pp. 160-171, 29 June 2001.

[24] B. Volckaerts, H. Ottevaere, P. Vynck, C. Debaes, P. Tuteleers, A. Hermanne, I. Veretennicoff, H. Thienpont, “Deep Lithography with Protons: a generic fabrication technology for refractive micro-optical components and modules”, Asian Journal of Physics, 2001, Vol.10, No.2, (Apr 2001), pp.195-214

[25] B. Volckaerts, H. Ottevaere, A. Vila, M. Muruzabal, C. Debaes, P. Vynck, P. Tuteleers, V. Baukens, A. Hermanne, I. Veretennicoff, H. Thienpont, “Deep Lithography with Protons: a generic technology for the fabrication of refractive micro-optical modules”, Proc. of the 2000 IEEE/LEOS International Conf. On Optical MEMS, Hawaii, pp. 104-105, 2000.

[26] E. W. Becker, W. Ehrfeld, R. Hagmann, A. Maner, D. Münchmeyer, “Fabrication of microstructures with high aspect ratios and great structural heights by synchrotron radiation lithography, galvanoforming and plastic moulding (LIGA proccess)”, Micro-electronic Eng., Vol. 4, pp. 35-56, 1986.

[27] H. Ottevaere, B. Volckaerts, J. Lamprecht, J. Schwider, A. Hermanne, I. Veretennicoff, H. Thienpont, “Two-dimensional plastic microlens arrays by deep lithography with protons: Fabrication and characterization.”, Journal of Optics A: Pure and Applied Optics; Vol.4, No.4 (July 2004), pp.S22-528

[28] P. Tuteleers, P. Vynck, H. Ottevaere, C; Debaes, A. Hermanne, I. Veretennicoff, H. Thienpont. “Replication of refractive micro-optomechanical components made with deep lithography with protons”, Proceedings of the SPIE, Sympsosium on Desgin, Test, Integration and Packaging of MEMS/MOEMS (DTIP2001), Vol. 4408, Cannes, France, 25-27 April 2001, pp. 329-337

[29] J. Van Campenhout, H. Van Marck, J. Depreitere and J. Dambre, “Opto-electronic FPGAs”, in IEEE Journal of Selected topics in Quantum Electronics, Vol. 5, No. 2 (march 1999), pp. 306-315.

[30] J. Van Campenhout, M. Brunfaut, W. Meeus, J. Dambre, and M. De Wilde, “Sense and nonsense of logic- level optical interconnect: reflections on an experiment," in Proceedings of the SPIE, SPIE Vol. 4455 (July 2001), pp. 151-159.