光互连CPU技术

Abstract

In the CMOS process, there are 4 materials, silicon nitride (Si3N4), silicon oxynitride (SiOxNy), silicon dioxide (SiO₂), and silicon (Si) that can be used as an optical waveguide. If silicon is not used, then the waveguide should be made with the remaining 3 materials. Among the methods that can inject light into those optical waveguides, the only possible method is to use a prism. It is demonstrated that an optical network chip can be realized by our three technologies and the cavity-type waveguide made out of silicon nitride and silicon dioxide.

The average propagation loss transmitted through the cavity-type waveguide, which was designed to minimize roughness on the waveguide side walls, was measured 0.258 dB/cm. Except for the reflection loss at the prism, the net coupling loss from VCSEL to the waveguide was measured 0.855 dB. The packaged size of VCSEL or photodiode has an area of 0.4 mm² and a height of 0.64 mm, which can be attached to the prism. The area for optical devices to be integrated on a 8 core CPU chip is much smaller, <30 mm² , than the available area, 174 mm² .

The prism used for injecting light into the waveguide can also be used as the optimal WDM filter in which two thin films with different refractive indices are alternately stacked on the bottom side. Theoretical calculations have proved the performance of a WDM filter. If anti-reflection films were coated on the two surfaces of the prism with materials, TiO₂ and Ta₂O₅, then the reflection loss at the prism was calculated to decrease from 2.75 dB to 1.11 dB for polyimide adhesive and from 2.73 dB to 0.52 dB for optical glue.

I. INTRODUCTION

A number of optical communication-based network  architectures for CPU application have been proposed, but they have not reached the stage of experimental  integration and testing on actual chips. Recently the lower  complexities of photonic processor chips and switching chips  integrated with electronic devices have been demonstrated  experimentally.  

The difficulties for implementing the proposed network  architectures can be listed as follows. The first is the  disadvantage of using SOI wafers. Wafers used in CPU and  memory chip production are ordinary bulk silicon. The fact that  electronic devices and optical devices can be manufactured  with the same CMOS process on the same wafer, which  motivated the research of silicon photonics, did not have many  benefits in reality. When using a silicon optical waveguide, the  optical waveguide and the optical device should be arranged in  the same silicon layer as the gates of the CPU. This is  disadvantageous not only in increasing the gate density of the  CPU, but also in designing metal layers of CPU or DRAM.  Contrary to the motives, for most proposed architectures, it is  assumed that optical devices and CPU were manufactured on  different wafers and bonded in a 3D package structure .

Second, the transmission loss of the silicon optical waveguide  is large. Silicon has a high refractive index (n = 3.45), which is  advantageous in making the size of the optical waveguide as  small as possible, but disadvantageous in designing a low  transmission loss of waveguide. Although ATAC assumed a  loss of 0.3 dB/cm, the minimum value reported in the paper  is 1~2 dB/cm, but in a typical CMOS process, it is  measured to be larger than 2 dB/cm.

Third, the yield and performance of optical modulators,  detectors, and ring resonators are far below implementation on  a network-on-chip. For the network to function, the failure rate  of each device needs to be lower than the redundant device rate  (generally 10 %). The average yield has to be over 90 %. When  manufacturing modulators and detectors by epitaxial growth of  Ge materials on silicon wafers, it is difficult to obtain a uniform  yield by the necessary ratio. Particularly, it becomes nearly  impossible when it is accompanied by fabrication-attributed  statistical errors such as the resonance wavelength of the ring  resonator. In this case, micro-heaters need to be  attached locally and heated for wavelength tuning. Heating  500,000~1.1 million resonators contained in a network chip  removes the advantage of optical signals. Furthermore,  it is assumed that the insertion loss of optical modulators is  approximately 1 dB for most proposed architectures, but  the reported insertion loss is about 3~6 dB, which is a  significant difference.

The fourth is the light source. An on-chip laser implemented  by joining an InP wafer to a silicon wafer suffers yield problems  similar to a Ge-based optical device. Most architectures are  presumed to use external light sources, but it is difficult  to integrate hundreds to thousands of lasers in the alignment of  optical fiber arrays with optical waveguide arrays on a chip. It  is also difficult to secure space to attach them on the side of the  chip when considering the cross-sectional area of the optical  fibers. The coupling loss measured when aligning one optical  fiber to a silicon optical waveguide is reported to be a minimum of 1 ~ 2 dB, but it is only accomplished by modifying  or adding a special structure at the inlet of the optical waveguide.  In general, coupling loss is measured to be 3~6 dB or more.

II. THREE TECHNOLOGIES  

First, an adhesive technology that is transparent, satisfies the  refractive index matching condition, maintains its adhesive  strength at 200-400 °C, and is not affected by solvents used in  the CMOS cleansing process such as BOE, isopropanol,  acetone, etc. needs to be achieved. There are no commercial  adhesives that are transparent with a refractive index greater  than 1.6 for the wavelength, 850-1550 nm. Adhesives that  maintain their strength in cleansing solvents and at 400 °C have  not been reported in any papers. Therefore, it is necessary to  develop a technology that can bond a prism to the surface of an  optical waveguide without using a commercial adhesive.

Second, a microlens that can produce collimated light under  the conditions of Fig. 1 needs to be achieved. Microlenses  consisting of quartz and silicon materials are commercially  available, but their refractive indices are not suitable for the current purpose. VCSELs fabricated with such microlenses in  references are not expected to be available  commercially in the near future. Third, a special package  technology for VCSEL and PD needs to be achieved. The  current commercial package has a hermetically sealed structure  that gets filled with nitrogen gas, maintaining a specific  distance between the laser light exit and the lens. This type of  packaging is overly bulky, and dozens to hundreds of VCSELs  and PDs cannot be integrated on the CPU chip. We have  developed an adhesive technology, a stand-alone polyimide  microlens, and a packaging technology that satisfies the  requirements.

Fig. 1. Conceptual diagram of optical coupling on SiON optical waveguide.

Fig. 2 (a) is a micrograph of a GaP prism bonded to the  surface of a SiON optical waveguide and (b) is an enlarged  photograph of the sideview. Thermo-set polyimide and our  specially developed procedure were used for bonding. The  adhesion characteristics between the prism and wafer are shown  in Table I. It had a transmissivity of over 99 % at the thickness  of 2 μm required for adhesion, and the refractive index was  measured to be up to 1.72 depending on the heat treatment  conditions applied to the polyimide. When force was applied  from the side to observe the adhesion strength, the prism broke  more times than it fell off.

Fig. 2. Images of prism bonded to SiON optical waveguide, (a) top view, (b)  sideview.

The radiation angle of VCSEL in the air is typically in the  range of 20~40°. The microlens must have the right value of  refractive index to collimate light emitted by the median value of 30°. Fig. 3 shows a scanning electron microscope (SEM)  image of the polyimide microlens array (a) and an enlarged  image (b) of one. The diameter of the microlens is 30 μm. The  properties of polyimide microlens are shown in Table II. As far  as we have studied, this is the first report of a stand-alone  polyimide microlens which was fabricated and imidized  without VCSEL.

Fig. 3. SEM images of microlens, (a) array, (b) single.

After attaching the microlens to the 25 Gbps VCSEL chip, the  shape of the output light was observed as shown in Fig. 4. The  picture was taken by placing an infrared card 7 cm vertically  from the VCSEL. Since the distance from the VCSEL to the  optical waveguide is <1 mm, the collimated light proves a  sufficient performance of the lens.

Fig. 4. Output light images of VCSEL before and after microlens attachment,  (a) before, (b) after.

III. MODELING WDM FILTER AND AR-COATING

A WDM filter with a verified performance among the  commercially available products is the thin film filter in which  two thin films with different refractive indices are alternately  stacked. Fig. 5 (a) shows the schematic diagram of four optical  wavelengths transmitted through a single optical waveguide by  coating WDM thin film filters on the bottom sides of the prisms.  For example, λ3 passes through the prism coated with the filter  of the corresponding wavelength and enters into the waveguide,  and it is reflected off the bottom sides of the prisms coated with  other wavelength filters and proceeds along the waveguide and  is emitted from the prism of the corresponding wavelength. Fig.  5 (b) graphically shows the structure of a WDM filter and the  light path.

(a)

(b)

Fig. 5. 4-channel WDM, (a) light input and output scheme, (b) thin film filter  and light path.

IV.CONCLUSION

Universities and corporations have proposed various network  architectures with optical communication technology, but they  have not reached a status of experimentally integrating and  testing them on actual chips. The difficulties can be attributed  in large part to the property of silicon used as the optical  waveguide.

In the CMOS process, there are 4 materials, silicon nitride  (Si3N4), silicon oxynitride (SiOxNy), silicon dioxide (SiO₂), and  silicon (Si) that can be used as an optical waveguide. If silicon  is not used, then the waveguide should be made with the  remaining 3 materials. Among the methods that can inject light  into those optical waveguides, the only possible method is to  use a prism. It is demonstrated that the difficulties can be  overcome by our three technologies and the cavity-type  waveguide made out of silicon nitride and silicon dioxide.

The average propagation loss transmitted through the cavitytype waveguide, which was designed to minimize roughness on  the waveguide side walls, was measured 0.258 dB/cm. Except  for the reflection loss at the prism, the net coupling loss from  VCSEL to the waveguide was measured 0.855 dB. The  packaged size of VCSEL or photodiode has an area of 0.4 mm² and a height of 0.64 mm, which can be attached to the prism.  The area for optical devices to be integrated on a 8 core CPU  chip is much smaller, <30 mm² , than the available area, 174  mm².

The prism used for injecting light into the waveguide can also  be used as the optimal WDM filter in which two thin films with  different refractive indices are alternately stacked on the bottom  side. Theoretical calculations have proved the performance of a  WDM filter. If anti-reflection films were coated on the two  surfaces of the prism with materials, TiO₂ and Ta₂O₅, then the  reflection loss at the prism was calculated to decrease from 2.75  dB to 1.11 dB for polyimide adhesive and from 2.73 dB to 0.52  dB for optical glue.

搜索“华林科纳行业观察”小程序查看完整内容，或者加微信13358064333/18106288187。