The static performance of the VCSEL is measured. Fig. 1(b) shows the power−current−voltage curves at 25 °C. The threshold current is 0.9 mA. By differentiating the voltage−current curve (dV/dI) at 9 mA, the VCSEL has a differential resistance of about 80 Ω due to the special design of the doping scheme. The spectrum of the VCSEL measured with an optical spectrum analyzer (AQ6370D) at 6 mA is shown in Fig. 1(c). Calculated from the spectrum, the VCSEL has a central wavelength of 857.31 nm and a root-mean-square (RMS) spectral width of 0.91 nm^[21]. The spectral width can be reduced by reducing the strength of the optical confinement. The size of the oxide aperture is estimated to be about 6 μm in diameter, deduced from the mode spacing between the fundamental mode and the first higher-order mode^[22].

The small signal frequency response (S21) of the full optical link with the VCSEL at different bias currents are measured with a vector network analyzer (Agilent PNA N5222A). It has a −3-dB bandwidth of 20.9, 21.8, 22.8, and 23.7 GHz at 6, 7, 8, and 9 mA (as shown in Fig. 1(d)), respectively. At 9 mA, a flat modulation response with a 1.3-dB peak is obtained, which is very beneficial to the large signal modulation^[15−17]. Large signal modulation experiments are performed in a back-to-back (BTB) configuration using a NRZ signal consisting of a 2³¹−1 bit long pseudorandom binary sequence (PRBS) and PAM-4 signal generated by a pattern generator (SHF 12103). The VCSEL is probed by a high-frequency probe in a ground-signal (GS) configuration. The light from the VCSEL is collected by a butt-coupled OM3 fiber into a high-speed photoreceiver (Thorlabs RXM25DF). The equivalent time sampling oscilloscope (Tektronix DSA 8300) is used to capture the eye diagrams. A peak-to-peak voltage (V_pp) swing of 650 mV is used to drive the VCSEL under a bias current of 9.5 mA without pre-emphasis. At the photoreceiver side, there is no equalization. Due to the limitation of the pattern generator, 35 Gbps NRZ modulation with open eyes at 25 °C is achieved, as shown in Fig. 1(e). 70 Gbps (35 Gbaud) PAM-4 eye diagrams are shown in Fig. 1(f). We can see clear eye openings with no pre-emphasis at the VCSEL side and no equalization at the receiver side.

Directly modulated 850-nm vertical-cavity surface-emitting lasers (VCSELs) with the advantages of low cost, high modulation speed, good reliability, and low power consumption, are the key sources in the optical interconnects with multimode fibers for the supercomputers, data centers, and machine learning applications^[1−3]. Typically, non-return-to-zero (NRZ) modulation format is used. The oxide-confined VCSEL with 71-Gbps NRZ modulation was realized with 130-nm BiCMOS IC with feed-forward equalization^[4]. Recently, a transition to pulse-amplitude modulation 4-level (PAM-4) modulation format has occurred to meet the higher data rate demand, evolving from 50 to 100 Gbps per lane, and beyond for 400G and 800G optical interconnect applications. In the past few years, 850-nm oxide-confined VCSELs with PAM-4 modulation have achieved many significant progresses, as shown in Table 1. Many achievements of VCSELs with PAM-4 modulation were with pre-emphasis, filter, and equalizer, which makes the design complex and increases the power consumption. In this letter we demonstrate the 850-nm oxide-confined VCSEL with 35-Gbps NRZ modulation and 70-Gbps PAM-4 modulation without equalization or digital signal processing (DSP).

In summary, we demonstrate the 850-nm oxide-confined VCSEL which achieves 35 Gbps NRZ modulation and 70 Gbps PAM-4 modulation without pre-emphasis, filter, or equalization for short-reach optical interconnects. Since above 24 GHz bandwidth and a flat small signal response are required for VCSELs for the 100 Gbps PAM-4 modulation with good eye diagram performance^{[15, 16, 23]}, we believe that our VCSELs can achieve 100 Gbps PAM-4 modulation with pre-emphasis, filter, and equalization. In future, we will optimize the VCSELs for a reduced spectral width, reduced relative intensity noise, higher bandwidth with a flat response, higher linearity, and reduced resistance.

The epi-structure of the VCSEL consists of 34-pair bottom distributed Bragg reflector (DBR), a half-wavelength cavity with multiple strained InGaAs quantum wells, and 21-pair top Al_0.12Ga_0.88As/Al_0.9Ga_0.1As DBR. The half-wavelength cavity is used to improve the carrier transport and enhance the confinement factor for a high relaxation resonance frequency^[19]. The cross-sectional view of the VCSEL structure is shown in Fig. 1(a). Multiple 30-nm-thick Al_0.96Ga_0.04As layers above the active region in the top DBR are used to reduce the parasitic capacitance. The positions of the Al_0.98Ga_0.02As layers above the active region are optimized to lower the differential resistance and reduce the strength of the transverse index guiding^[20]. The 34-pair bottom DBR includes AlAs/Al_0.12Ga_0.88As pairs with a high thermal conductivity to facilitate effective heat removal for a low thermal impedance. The DBRs have graded interfaces and modulation doping schemes for low loss and low resistance. In the top and bottom DBRs, the average doping level is reduced towards the cavity to minimize the loss. The oxide apertures are formed by the wet oxidation of 30-nm-thick Al_0.98Ga_0.02As layers for the current and optical confinements. VCSELs are fabricated with processing flow for oxide-confined GaAs-based VCSELs. VCSELs have mesas of 24-μm in diameter, and oxide apertures of about 6-μm in diameter for the current and optical confinements. Benzocyclobutene (BCB) is used under the coplanar pads to reduce the parasitic capacitance. At last, the wafer is thinned to a thickness of 150 μm.