CMOS-compatible on-chip optical interconnects
Date
2009
Journal Title
Journal ISSN
Volume Title
Abstract
The increase in complexity of integrated circuits (ICs) over the past five decades has resulted increasing demands on the interconnect layers. In the past decade, the ability of conventional "electrical signal down a metal wire" interconnect to keep up with the increasing demands placed on interconnect has come more and more into question. To meet the increasing demands on interconnect and to get around the limitation of conventional "metal wire" interconnect, various forms of optical interconnect have been proposed.
In the Optoelectronics Research Group at Colorado State University, we have conceived, developed and investigated a proposal to place a nanoscale fiber optic network on an integrated circuit. In contrast to other proposals which use materials and processing which are arguably not part of a standard CMOS process and thus raise equipment compatibility or contamination concerns, our proposal uses materials that exist in any reasonably standard CMOS process. This compatibility greatly reduces barriers to adoption. Waveguide cores are made from silicon nitride (a standard dielectric film), while the waveguide cladding is made of various types of silicon dioxide (another standard dielectric film). Photodetectors in the system are made from polysilicon, a standard gate material. Two generations of optical interconnect test chips were fabricated in a standard CMOS process at a commercial fabrication facility with a few minor variations to meet the needs of the test chip.
Following fabrication of each generation of the optical interconnect test chip, the devices on the test chip were characterized for both DC and AC performance. The waveguide loss, which was expected to be between 0.5 dB/cm and 2 dB/cm, was found to be much higher than expected, approximately 8.5 dB/cm. This value was essentially independent of waveguide width from 0.5 µm to 8 µm. This high waveguide loss was attributed to an unexpected absorption mechanism in the phosphosilicate glass (PSG) form of silicon dioxide which made up part of the lower cladding.
Detector dark current was extremely low, typically 5 nA or less at a 5 V bias for all device sizes. Dark current was found to follow the hyperbolic sine I-V curve that is characteristic of undoped polysilicon due to thermionic emission over the polysilicon grain boundaries. First generation DC responsivity was measured as high as 0.35 A/W at 654 nm, while second generation DC responsivity measured as high as 1.3 A/W at 690 nm. The higher responsivity observed on second generation material is attributed to the additional thermal processing seen by second generation material. Responsivity increased at smaller contact spacing, however no sign of saturation in responsivity was observed at the smallest contact spacing in this work, 2.14 µm. This suggests that devices with even smaller contact spacing should have even higher responsivity. The dependence of responsivity ℜ = ℜ0(KµτV/S2) on voltage V and contact spacing S, suggested that the detectors worked as photoconductors rather than photodiodes. µ is the carrier mobility, τ is the carrier lifetime, and K depends on the ratio of bulk recombination to contact recombination. Further suggesting a photoconductive, rather than photodiode, principle behind the response was the fact that the photocurrent in the symmetric detector structure increased with voltage, showing no signs of saturation all the way to device breakdown.
AC testing of second generation devices measured pulses with full-width half-max (FWHM) as low as 0.89 ns and 10%-90% rise times as low as 0.39 ns. The values are equipment-limited, as we were unable to generate Gaussian optical pulses with FWHM much less than 600 ps. In addition, the transimpedance amplifier (TIA) used to convert the photocurrent signal into a voltage had a 1 GHz bandwidth, which limits the measureable rise time. The shape of the detector photocurrent response to a Gaussian optical input indicated that more than one recombination mechanism was present. An extremely good fit to the data was found by combining a bulk recombination process that dominated at lower optical power levels with a contact recombination process that became dominant at higher optical power levels.
DC responsivity and AC pulse width were both affected by a photoconductive gain process. This process allowed more than one carrier to be counted in the external circuit for every absorbed photon, resulting in quantum efficiencies as high as 234%. However, the same photoconductive gain process effectively allowed carriers to live longer than the transit time between the contacts, causing an exponential decay process to be observed in the AC pulse response and increasing the FWHM of the output pulses. The gain process is due to barrier lowering at the Schottky contacts, enhanced by capture of photogenerated charges in states physically near the contacts.
In the Optoelectronics Research Group at Colorado State University, we have conceived, developed and investigated a proposal to place a nanoscale fiber optic network on an integrated circuit. In contrast to other proposals which use materials and processing which are arguably not part of a standard CMOS process and thus raise equipment compatibility or contamination concerns, our proposal uses materials that exist in any reasonably standard CMOS process. This compatibility greatly reduces barriers to adoption. Waveguide cores are made from silicon nitride (a standard dielectric film), while the waveguide cladding is made of various types of silicon dioxide (another standard dielectric film). Photodetectors in the system are made from polysilicon, a standard gate material. Two generations of optical interconnect test chips were fabricated in a standard CMOS process at a commercial fabrication facility with a few minor variations to meet the needs of the test chip.
Following fabrication of each generation of the optical interconnect test chip, the devices on the test chip were characterized for both DC and AC performance. The waveguide loss, which was expected to be between 0.5 dB/cm and 2 dB/cm, was found to be much higher than expected, approximately 8.5 dB/cm. This value was essentially independent of waveguide width from 0.5 µm to 8 µm. This high waveguide loss was attributed to an unexpected absorption mechanism in the phosphosilicate glass (PSG) form of silicon dioxide which made up part of the lower cladding.
Detector dark current was extremely low, typically 5 nA or less at a 5 V bias for all device sizes. Dark current was found to follow the hyperbolic sine I-V curve that is characteristic of undoped polysilicon due to thermionic emission over the polysilicon grain boundaries. First generation DC responsivity was measured as high as 0.35 A/W at 654 nm, while second generation DC responsivity measured as high as 1.3 A/W at 690 nm. The higher responsivity observed on second generation material is attributed to the additional thermal processing seen by second generation material. Responsivity increased at smaller contact spacing, however no sign of saturation in responsivity was observed at the smallest contact spacing in this work, 2.14 µm. This suggests that devices with even smaller contact spacing should have even higher responsivity. The dependence of responsivity ℜ = ℜ0(KµτV/S2) on voltage V and contact spacing S, suggested that the detectors worked as photoconductors rather than photodiodes. µ is the carrier mobility, τ is the carrier lifetime, and K depends on the ratio of bulk recombination to contact recombination. Further suggesting a photoconductive, rather than photodiode, principle behind the response was the fact that the photocurrent in the symmetric detector structure increased with voltage, showing no signs of saturation all the way to device breakdown.
AC testing of second generation devices measured pulses with full-width half-max (FWHM) as low as 0.89 ns and 10%-90% rise times as low as 0.39 ns. The values are equipment-limited, as we were unable to generate Gaussian optical pulses with FWHM much less than 600 ps. In addition, the transimpedance amplifier (TIA) used to convert the photocurrent signal into a voltage had a 1 GHz bandwidth, which limits the measureable rise time. The shape of the detector photocurrent response to a Gaussian optical input indicated that more than one recombination mechanism was present. An extremely good fit to the data was found by combining a bulk recombination process that dominated at lower optical power levels with a contact recombination process that became dominant at higher optical power levels.
DC responsivity and AC pulse width were both affected by a photoconductive gain process. This process allowed more than one carrier to be counted in the external circuit for every absorbed photon, resulting in quantum efficiencies as high as 234%. However, the same photoconductive gain process effectively allowed carriers to live longer than the transit time between the contacts, causing an exponential decay process to be observed in the AC pulse response and increasing the FWHM of the output pulses. The gain process is due to barrier lowering at the Schottky contacts, enhanced by capture of photogenerated charges in states physically near the contacts.
Description
Rights Access
Subject
dark current
leaky mode waveguides
optical interconnects
photoconductivity
electrical engineering