Krzysztof Szczerba

Optical interconnects (OIs) based on vertical-cavity surface-emitting lasers (VCSELs) are the main workhorse within data centers, supercomputers, and even vehicles, providing low-cost, high-rate connectivity. VCSELs must operate under extremely harsh and time-varying conditions, thus requiring adaptive and flexible designs of the communication chain. Such designs can be built based on mathematical models (model-based design) or learned from data (machine learning (ML) based design). Various ML techniques have recently come to the forefront, replacing individual components in the transmitters and receivers with deep neural networks. Beyond such component-wise learning, end-to-end (E2E) autoencoder approaches can reach the ultimate performance through co-optimizing entire parameterized transmitters and receivers. This tutorial paper aims to provide an overview of ML for VCSEL-based OIs, with a focus on E2E approaches, dealing specifically with the unique challenges facing VCSELs, such as the wide temperature variations and complex models.

Optical interconnects (OIs) based on vertical-cavity surface-emitting lasers (VCSELs) are the main workhorse within data centers, supercomputers, and even vehicles, providing low-cost, high-rate connectivity. VCSELs must operate under extremely harsh and time-varying conditions, thus requiring adaptive and flexible designs of the communication chain. Such designs can be built based on mathematical models (model-based design) or learned from data (machine learning (ML) based design). Various ML techniques have recently come to the forefront, replacing individual components in the transmitters and receivers with deep neural networks. Beyond such component-wise learning, end-to-end (E2E) autoencoder approaches can reach the ultimate performance through co-optimizing entire parameterized transmitters and receivers. This tutorial paper aims to provide an overview of ML for VCSEL-based OIs, with a focus on E2E approaches, dealing specifically with the unique challenges facing VCSELs, such as the wide temperature variations and complex models.

Transition from on-off keying to 4-level pulse amplitude modulation (PAM) in VCSEL based optical interconnects allows for an increase of data rates, at the cost of 4.8 dB sensitivity penalty. The resulting strained link budget creates a need for accurate VCSEL models for driver integrated circuit (IC) design and system level simulations. Rate equation based equivalent circuit models are convenient for the IC design, but system level analysis requires computationally efficient closed form behavioral models based Volterra series and neural networks. In this paper we present and compare these models.

Since energy dissipation and latency in optical interconnects are of utmost concern, such links are often operated without forward error correction. We propose a low-complexity noniterative two-error correcting Bose-Chaudhuri-Hocquenghem (BCH) decoder circuit that significantly relaxes the stringent optical modulation amplitude requirements on the transmitter, which help to reduce laser and laser driver energy dissipation. We demonstrate, in a 28-nm CMOS process technology, that the introduction of this BCH circuit in an uncoded link leads to energy-per-bit reductions of 25%, even when the impact of code rate on receiver energy dissipation is considered.

In this paper, we present a simple circuit-level VCSEL model working over a broad range of ambient temperatures, while accounting for self-heating effects. We present the static, dynamic small signal and dynamic large signal performance compared to measurement results.

We present an investigation into the effects of photon lifetime and the associated damping of the modulation response on 50 Gbps 4-pulse-amplitude modulation (4-PAM) signal generation with directly modulated 25G class vertical-cavity surface-emitting lasers (VCSELs). While 4-PAM benefits from higher slope efficiency and output power, it is shown that the greater impact of intensity noise and dynamic nonlinearities implies that a VCSEL with a longer photon lifetime and more damped response is needed when transitioning from existing 25 Gbps OOK to 50 Gbps 4-PAM using the 25G VCSEL technology.

Multilevel pulse amplitude modulation (M-PAM) is gaining momentum for high-capacity and power-efficient cloud computing. Compared to the classic on-off keying (OOK) modulation, high-order PAM yields better spectral efficiency but is also more susceptible to physical layer degradation effects. We develop a cross-layer analysis framework to examine the PAM transmission performance in data center network environments supporting both optical multicasting and wavelength routing. Our analysis is conducted on a switch architecture based on an arrayed-waveguide grating (AWG) core and distributed broadcast domains, exhibiting different physical paths, and random, uncontrolled crosstalk noise. Reed-Solomon coding with rate adaptation is incorporated into PAM transceivers to compensate for impairments. Our Monte Carlo simulations point to the significant impact of AWG crosstalk on higher order PAM in wavelength-reuse architectures and the importance of code rate adaptation for signals traversing multiple routing stages. According to our study, 8-PAM offers the highest effective bit rates for signals terminating in one broadcast domain and performs poorly when considering interdomain connectivity. On the other hand, the impairment-induced degradation of interdomain capacity for 4-PAM can be limited to 20.7%, making it better suited for connections spanning two broadcast domains and a crosstalk-rich stage. Our results call for software-defined PAM transceiver designs in support of both modulation order and code rate adaptation.

We present our recent work on high-speed optical interconnects with advanced modulation formats and directly modulated 850 nm VCSELs. Data transmission at nearly 100 Gbps was achieved with 4-PAM. Forward error correction, equalization and preemphasis are also explored. The system aspects of the advanced modulation formats and their impact on the VCSEL requirements are discussed. Requirements on the optical output power, frequency response and the relative intensity noise are discussed. Finally, co-optimization of the VCSELs and VCSEL driver amplifiers in CMOS and InP technologies is discussed.