Introduction

Avalanche photodetectors (APD) are highly valued for their high bandwidth, high gain, and compact integration, making them widely applicable in optical communications, optical interconnects, light detection and ranging (LiDAR) systems, quantum computing, and biosensing1,2,3. As artificial intelligence (AI), machine learning, and cloud computing continue to expand rapidly, the demand for computational power and high-speed data transmission in large-scale data centers and high-performance computing (HPC) is growing exponentially4,5,6. This surge has intensified the need for transceivers with higher bandwidth density. However, current optical link and power budgets are struggling to keep pace with these escalating demands4. As a pivotal component in optical receivers, APDs play a crucial role in enhancing link performance by providing internal gain, which helps reduce power consumption and optimize signal integrity. These characteristics are essential for advancing optical receiver systems toward higher bandwidth and lower power operation7,8.

Gain bandwidth product (GBP) is a key metric in evaluating APD performance in high-speed optical communication and is strongly influenced by the impact ionization coefficient ratio. While most III-V materials exhibit an impact ionization coefficient ratio in the range of k ≈ 0.1–19,10, silicon (Si) stands out with an exceptionally low ratio of k ≈ 0.0211,12. However, Si’s cutoff wavelength of 1100 nm limits efficient photo-carrier generation in fiber communication wavelengths, necessitating alternative materials. Germanium (Ge) is a preferred material for silicon-based PDs due to its superior absorption efficiency at fiber communication wavelengths and higher carrier mobility compared to Si. However, the impact ionization coefficient ratio of Ge (k ≈ 1)13,14 is significantly higher, potentially leading to increased noise. To leverage the high absorption of Ge while benefiting from the low ionization coefficient ratio of Si, researchers have proposed APDs with separate absorption charge multiplication (SACM) structures15. Fabricated on a standard silicon-on-insulator (SOI) platform, the Si-Ge APD offers higher gain, excellent sensitivity, and superior integration density, making it an optimal choice for high-performance optical receivers16,17.

In current SACM structure research, vertical SACM-type APDs require metal electrode deposition on the Ge surface18,19. However, this approach induces Fermi-level pinning, degrading the ohmic contact and increasing contact resistance. Additionally, the metal presence on the Ge surface introduces more dangling bonds, leading to higher leakage current, which in turn reduces device responsivity and increases noise. Moreover, metal contacts facilitate additional photogenerated carrier recombination, further compromising optoelectronic performance. In contrast, lateral SACM Si-Ge APDs mitigate these challenges while also simplifying fabrication. Dai et al. proposed a series of lateral SACM Si-Ge APDs, demonstrating a significant GBP improvement from 615 GHz20 to 1440 GHz21, and later to 3036 GHz22. This progression has provided valuable insights into GBP enhancement strategies. However, this design still exhibits relatively high dark current at the operating voltage, which can hinder the achievement of higher sensitivity and gain. A recent study reported an improved lateral SACM Si-Ge APD with a bandwidth of 53 GHz and a GBP of 1033 GHz23. The high bandwidth was achieved by incorporating a large inductor ( ~ 700 pH) into the APD electrode through inductive peaking. However, in the SACM structure, the multiplication gain and GBP remain constrained by the Ge-Si doping contact due to the high dark current resulting from significant tunneling currents at high voltages. Additionally, this approach may introduce challenges in impedance matching with the transimpedance amplifier (TIA) and signal transmission. Currently, researchers continue exploring various strategies to further enhance APD GBP for higher-speed and higher-gain detection24. However, achieving a significantly higher GBP remains a critical and challenging research direction.

In this work, we present a high-speed Si-Ge APD with a record-breaking GBP of 7564 GHz, which exceeds previously reported values. Utilizing a lateral SACM structure, we design an APD that spatially separates the Ge absorption region from the doped regions. By minimizing the electric field in Ge while maintaining efficient carrier transport and multiplication, the dark current is significantly reduced. To further enhance device responsivity, we incorporate a tapered waveguide at the front end and a distributed Bragg reflector (DBR) at the rear. Experimental results show that under a reverse bias voltage of 12.5 V, the APD exhibits a dark current of 8.4 μA and achieves a maximum GBP of 7564 GHz at an input optical power of −24 dBm, while maintaining a low effective k value of 0.05. We evaluate the APD performance for high-speed signals and obtain clear eye diagrams for 64 Gbps and 100 Gbps non-return-to-zero (NRZ) signals, as well as 128 Gbps and 200 Gbps four-level pulse amplitude modulation (PAM4) signals. The APD also exhibits high sensitivity, achieving −20.8 dBm and −17.1 dBm for 64 Gbps and 100 Gbps NRZ signals, respectively, both of which meet the hard-decision forward error correction (HD-FEC) threshold of BER = 3.8 × 10−3. For 128 Gbps and 200 Gbps PAM4 signals, the sensitivities are −14.2 dBm and −9.6 dBm, respectively, both within the HD-FEC limit. Furthermore, we perform high-speed eye diagram measurements for 200 Gbps PAM4 signals across eight APDs, each tuned to the specific wavelengths required for eight-channel dense wavelength division multiplexing (DWDM). This manuscript represents a substantial extension of our previous OFC work25, expanding the study from a single-channel device to an eight-channel APD array, and providing detailed design and analysis, along with a further systematic investigation of excess-noise characteristics. These remarkable results highlight the immense potential of this APD for applications in optical communications, quantum technology, LiDAR, and beyond.

Results

Device structure and design

Figure 1a presents a three-dimensional (3D) schematic of the lateral p + -i-p-n+ doped SACM Si-Ge APD, incorporating a DBR at the rear to enhance device responsivity. The design process begins with optimizing the electric field distribution to achieve high performance. Figure 1b illustrates the cross-sectional doping profile of the Si-Ge APD, featuring a non-contact design between the Ge region and the Si-doped areas to effectively control the electric field within the Ge. To quantitatively describe this non-contact design strategy, we optimize the doping positions on the rib waveguide, denoted as Wp and ΔW, respectively. Additionally, thicker metal interconnects are employed to further enhance device performance. The Ge absorption region dimensions are 9.1 μm (length) × 0.5 μm (width) × 0.26 μm (height). Given Ge’s high impact ionization coefficient, optimization efforts focus on the doping regions adjacent to the Ge layer, specifically adjusting the widths of Wp and ΔW. The primary goal is to maximize ionization within Si while suppressing ionization in Ge, requiring the electric field strength in Si to exceed 3 × 10⁷ V/m, while ensuring it remains below 2 × 10⁷ V/m in Ge26,27. Additionally, to facilitate efficient carrier transport at saturation velocity within Ge, the electric field in Ge is designed to exceed 1 × 10⁶ V/m28.

Fig. 1: Device schematic and simulation results of the Si-Ge APD.
Fig. 1: Device schematic and simulation results of the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

a Three-dimensional schematic of the Si-Ge APD. b Cross-sectional view of the doping distribution within the Si-Ge APD, with an inset showing the simulated electric field distribution under a reverse bias of 12.5 V. c Simulated electric field distribution along the black cutline in (b) for APDs with varying widths of p-type Si region (Wp = 150 nm, 175 nm, 200 nm, 225 nm, and 250 nm) under a reverse bias of 12.5 V. d Simulated electric field distribution along the black cutline in (b) for APDs with varying widths of the p+ type Si region (ΔW = 70 nm, 90 nm, 110 nm, 130 nm, and 150 nm) under a reverse bias of 12.5 V. e Reflectivity of the DBR design as a function of the number of periods. The inset illustrates the reflectivity spectrum for DBRs with 2, 3, and 5 periods.

The electric field design begins with optimizing the multiplication region doping width Wp to maintain a high electric field within the multiplication region while preventing its extension into the absorption region. Figure 1c depicts the electric field distribution along the black cutline of the ridge waveguide in Fig. 1b under varying Wp conditions. As the Wp increases, the electric field strength in the multiplication region rises, but the high electric field region gradually shifts toward the absorption region, potentially increasing the electric field in Ge and degrading device performance. To balance these effects, Wp is set at 225 nm. Next, ΔW is optimized to confine the electric field effectively within the charge region. Figure 1d illustrates the electric field distribution along the black cutline of the ridge waveguide in Fig. 1b under different p+ doping width conditions, with an inset providing a magnified view of the localized region. If the doping position is too far from the Ge layer, the electric field is not fully confined within the charge region, potentially inducing additional current in the Si-doped area and increasing the device’s shot noise. Based on the above considerations, i.e., to ensure effective confinement of the optical field, ΔW is set to 110 nm.

To further enhance performance, graded doping is employed to reduce resistance and capacitance. The p-type doping concentrations are set at 7.1 × 1019 cm−3 (p + +), 5.3 × 1018 cm−3 (p + ), and 2.1 × 1017 cm−3 (p), while the n-type doping concentrations are set at 1.0 × 1020 cm−3 (n + +) and 1.0 × 1019 cm−3 (n + ), respectively. Additionally, a lateral structure is adopted to minimize light absorption by metal electrodes and reduce dark current caused by defects near the heterojunction under breakdown voltage conditions. The inset in Fig. 1b presents the simulated cross-sectional electric field results under a reverse bias of 12.5 V, confirming effective electric field confinement. Through optimized doping position and concentration, together with precise control of Wp and ΔW enabled by the non-contact design, the high electric field is effectively confined within the Si multiplication region, while the Ge absorption region is maintained at a relatively low electric field, thereby achieving precise electric-field engineering. Under this field configuration, avalanche multiplication is dominated by an electron-driven quasi-unipolar mechanism, in which impact ionization in Si is maximized whereas ionization in Ge is strongly suppressed. As a result, the excess noise is effectively reduced and the overall device performance is significantly enhanced.

The APD couples O-band optical signals from the Si waveguide into Ge. To mitigate coupling loss caused by abrupt width changes between the Si and Ge waveguides, a tapered waveguide is employed. Traditionally, increasing Ge volume has been a primary method for enhancing responsivity. However, as the volume of Ge increases, the carrier transit time is prolonged, reducing the response speed and adversely impacting the device’s electrical performance. To address this issue, we integrate a DBR at the rear and incorporated a tapered waveguide at the front to confine more light within the absorption region, thereby enabling secondary light absorption and improving overall responsivity.

The DBR structure in this device is a first-order DBR, chosen for its low scattering losses and high reflectivity. As depicted in Fig. 1e, when the number of grating periods reaches five, the reflectivity exceeds 90%, effectively reflecting light at the tail. The inset in Fig. 1e illustrates that, with five grating periods, the operational wavelength range extends from 1100 nm to 1700 nm, perfectly covering the O-band. Figure 2 presents 3D finite-difference time-domain (FDTD) simulation results for two APD structures. The evanescently coupled optical field distributions for APDs without and with a DBR are shown in Fig. 2a, c, respectively. In the absence of a DBR, significant light leakage occurs at the end of the APD. In contrast, incorporating a DBR significantly reduces this leakage by reflecting the unabsorbed light back into the Ge region, thereby enhancing responsivity. To fully utilize this reflective effect, the Ge length is optimized to 9.1 μm (see Supplementary Note 1), rather than 9.0 μm or 9.2 μm, as this specific length allows more unabsorbed light to be reflected and reabsorbed during a second pass. However, standing waves are observed in the field distribution, mainly due to mode mismatch between the Si waveguide and Ge layer as well as reflections from the rear DBR. Simulations indicate that the optical return loss (ORL) of this device is −22.5 dB (see Supplementary Note 2). To further verify the effect of the DBR, top-view field distributions in Fig. 2b, d show that the inclusion of the DBR significantly increases the optical intensity in the Ge absorption region. The inclusion of a five-period DBR at the rear enhances the overall light confinement within the Ge layer. Additionally, a front-end tapered waveguide is employed to minimize input reflection losses. As a result, the device’s responsivity improves by 35%.

Fig. 2: 3D-FDTD simulation diagram of the Si-Ge APD.
Fig. 2: 3D-FDTD simulation diagram of the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

Cross-sectional view of optical field distributions: (a) without DBR and (c) with DBR. Top-view of optical field distributions in the Ge absorption region: (b) without DBR and (d) with DBR.

Static measurements

Figure 3a shows the I-V characteristics of the fabricated Si-Ge APD under various bias voltages, both with and without incident light, with optical power ranging from −26 dBm to −12 dBm. The grating couplers (GCs) used in this work are characterized using a test structure consisting of a pair of GCs connected by a short straight waveguide. The measured insertion loss is 6.0 dB/facet at 1310 nm. At a reverse bias of 3.5 V, the APD exhibits an optical internal responsivity of 0.75 A/W at unit gain, determined by comparison with a reference p-i-n PD fabricated on the same chip and sharing identical structural parameters (see Supplementary Note 3), including Ge volume, waveguide dimensions, and GC design2,20,22,29. The responsivity of our designed APD at 1310 nm is lower than the theoretical value (1.06 A/W) because, to accommodate the bandwidth limitations imposed by a large Ge volume, part of the responsivity is intentionally sacrificed. The dark current initially increases gradually and then rises sharply as the voltage approaches breakdown. At a reverse bias voltage of 12.5 V, the dark current remains low at 8.4 μA, allowing the device to achieve exceptionally high multiplication gain. The breakdown voltage of the Si-Ge APD is 13 V, corresponding to a dark current of 100 μA20. High multiplication gain enables high-sensitivity data reception in optical systems. Figure 3b illustrates the multiplication gain at reverse bias voltages ranging from 9 V to 13.5 V for incident optical power levels between −26 dBm and −12 dBm. The avalanche multiplication gain G(V) is defined as the difference between the photocurrent (Ip) and the dark current (Id) at a given bias voltage, normalized to the corresponding value at the unity gain point V0: \({{\rm{G}}}\left({{\rm{V}}}\right)=({{\rm{Ip}}}({{\rm{V}}})-{{\rm{Id}}}({{\rm{V}}}))/({{\rm{I}}}{{\rm{p}}}({\rm{V}_{0}})-{{\rm{I}}}{{\rm{d}}}({\rm{V}_{0}}))\)23. The gain increases monotonically as the incident optical power decreases, reaching its maximum at a reverse bias of 12.5 V. However, with further increases in reverse bias, the gain starts to decrease due to the dark current rising more rapidly than the photocurrent.

Fig. 3: Performance characterization of the Si-Ge APD.
Fig. 3: Performance characterization of the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

a Measured IV curves with different input optical powers. b The dependence of gain on input optical power at various reverse bias voltages. c Measured OE responses of the APD at various reverse bias voltages under an input optical power of −24 dBm. The inset shows the 3-dB OE bandwidth of the APD for input optical powers ranging from −26 dBm to −12 dBm at a reverse bias voltage of 12.5 V. d Measured GBP of the APD under reverse bias voltages of 9 V, 10 V, 11 V, 12 V, and 12.5 V, as well as at different input optical powers. The inset shows the relationship between bandwidth and gain of the APD at an input optical power of −24 dBm. e Measured excess noise factor versus multiplication gain for the APD.

Small-signal measurements

To characterize the optoelectronic (OE) bandwidth of the APD through small-signal measurements, the test setup is illustrated in Fig. 4a. The optical signal emitted by the O-band laser is first directed into a lightwave component analyzer (LCA), then sequentially passes through a handheld variable optical attenuator (VOA) and a polarization controller (PC), before being vertically coupled into the device under test (DUT) via a GC. Figure 4c, d shows the physical photograph and a cross-sectional scanning electron microscope (SEM) image of the APD, respectively. The photocurrent generated by the APD is routed back to the LCA through a bias-tee to measure the bandwidth. As shown in Fig. 3c, when the input optical power is −24 dBm, the APD exhibits a 3-dB OE bandwidth of 8 GHz at a reverse bias of 7 V, primarily due to incomplete depletion of the Ge absorption region. As the reverse bias voltage increases, the 3 dB OE bandwidth gradually increases, reaching a maximum of 31 GHz at a reverse bias voltage of 12.5 V. However, with further increases in reverse bias, the 3 dB OE bandwidth starts to decrease, primarily due to the dominance of the avalanche buildup time in the frequency response of the APD. The inset of Fig. 3c shows the variation of the 3-dB OE bandwidth under input optical powers ranging from −26 dBm to −12 dBm at a fixed reverse bias voltage of 12.5 V. As the input optical power decreases, the bandwidth gradually reduces. At −26 dBm, the bandwidth is 30 GHz, representing a 16.6% reduction compared to −12 dBm. This behavior arises because, at lower input optical powers, the generation and collection of photogenerated carriers slow down, limiting the APD response speed and consequently reducing the 3-dB bandwidth. Figure 3d illustrates the multiplication gain and GBP characteristics at reverse bias voltages of 9 V, 10 V, 11 V, 12 V, and 12.5 V, corresponding to five different multiplication gain points for input optical power levels ranging from −12 dBm to −24 dBm. At an input power of -16 dBm and a reverse bias of 12.5 V, the APD achieves a multiplication gain of 73 and a 3 dB OE bandwidth of 36 GHz, corresponding to a maximum GBP of 2628 GHz. When the input power is reduced to −24 dBm at the same 12.5 V reverse bias, the GBP of the APD increases to 7564 GHz. The inset of Fig. 3d shows the trends of gain and bandwidth at an input optical power of −24 dBm. Both metrics increase with bias voltage until peaking, after which they decline. This degradation is attibuted to excessive noise and prolonged avalanche build-up time at high voltages, which deteriorate the APD’s performance.

Fig. 4: Experimental setup for characterizing the Si-Ge APD.
Fig. 4: Experimental setup for characterizing the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

a O-E bandwidth test. b Eye-diagram test. c Photograph of the DUT. d Cross-sectional SEM image of the DUT.

Noise measurement

For APDs, the sensitivity and signal-to-noise ratio primarily benefit from multiplication gain, although the shot noise current is also amplified. To characterize and analyze the device noise, we employed a noise analyzer. The shot noise current is given by: \({{{\rm{I}}}}_{\left\{{{\rm{shot}}}\right\}}^{2}=2{{\rm{q}}}{{{\rm{M}}}}^{2}\left.\left({{{\rm{I}}}}_{{{\rm{photo}}}}+{{{\rm{I}}}}_{{{\rm{dark}}}}\right)\right){{\rm{F}}}\left({{\rm{M}}}\right)\Delta {{\rm{f}}}\), where q denotes the electron charge, M is the multiplication gain, F(M) represents the excess noise factor, Δf is the operating bandwidth, Iphoto is the photocurrent at unity gain, and Idark is the dark current at unity gain. The excess noise factor of an APD can be expressed as: \({{\rm{F}}}\left({{\rm{M}}}\right)={{\rm{k}}}\cdot {{\rm{M}}}+\left(1-{{\rm{k}}}\right)\cdot (2-1/{{\rm{M}}})\), where \({{\rm{k}}}={{\rm{\beta }}}/{{\rm{\alpha }}}\) is the impact ionization coefficient ratio, with α and β being the electron and hole ionization coefficients in the multiplication region. To evaluate the APD excess noise, we measured and plotted F(M) as a function of multiplication gain M. The excess-noise factor F and the corresponding effective k value were extracted at low multiplication gain, where the fitting is more reliable and less affected by measurement-chain noise. For comparison, theoretical excess noise factors are calculated using the local-field model for k values in the range of 0.03-0.20, as shown in Fig. 3e30. Although the GBP is evaluated at higher multiplication gain, prior studies indicate that for APDs employing thin multiplication regions and electron-dominated avalanche multiplication, the effective k factor does not vary significantly with the multiplication gain M31,32,33,34,35. At 1310 nm, the extracted effective k value is approximately 0.05, which is close to that of Si (k ≈ 0.02), confirming that avalanche multiplication predominantly occurs in the Si region of our APD. This value is also significantly lower than that of Ge APDs (k ≈ 0.2–0.5) and conventional III-V APDs (k ≈ 0.1–1)2,36,37,38,39. Previous studies have shown that as the k value decreases, the GBP of APDs increases2,5,13,40. This clear trend indicates that a lower k value effectively suppresses excess noise under high multiplication gain, which is critical for achieving higher GBP in APDs41.

Eye Diagram and Bit Error Rate Measurements

To assess the signal reception quality, an eye diagram measurement was conducted. Figure 4b shows the experimental setup for measuring the APD eye diagram and bit error rate (BER). An O-band tunable laser was used as the optical source. The optical signal first passed through a PC and was then modulated by a thin-film lithium niobate (TFLN) Mach–Zehnder modulator (MZM), which was driven by an amplified high-speed electrical signal. After a second PC for polarization adjustment, the modulated optical signal was vertically coupled into the DUT via a GC. The photocurrent generated by the APD was directed through a bias-tee to an oscilloscope for eye diagram and TDECQ analysis, and then sent to a BER tester to assess the APD’s sensitivity.

The eye diagram measurements were conducted with an input optical power of −16 dBm and a reverse bias voltage of 12.5 V, with no TIA connected to the APD. Figure 5a, b presents the eye diagrams measured at data rates of 64 Gbps NRZ and 100 Gbps NRZ, with signal-to-noise ratios (SNRs) of 8.26 dB and 6.08 dB, respectively. Figure 5c, d shows the eye diagrams for data rates of 128 Gbps PAM4 and 200 Gbps PAM4, with PAM4 transmitter and dispersion eye closure quaternary (TDECQ) values of 1.88 dB and 3.23 dB, respectively.

Fig. 5: High-speed performance of the Si-Ge APD.
Fig. 5: High-speed performance of the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

Eye diagrams measured at an input optical power of −16 dBm for data rates of (a) 64 Gbps NRZ, (b) 100 Gbps NRZ, (c) 128 Gbps PAM4, and (d) 200 Gbps PAM4. e Measured BERs at a reverse bias voltage of 12.5 V for NRZ and PAM4 signals.

To evaluate the sensitivity of the APD, BER tests were conducted. The upper partion of Fig. 5e shows the BER performance for NRZ signals at 64 Gbps and 100 Gbps with a reverse bias voltage of 12.5 V. The corresponding sensitivities at the HD-FEC threshold were measured as −20.8 dBm and −17.1 dBm, respectively. The lower partion of Fig. 5e illustrates the sensitivities for PAM4 signals at 100 Gbps and 200 Gbps under the same bias voltage, with sensitivies of −14.2 dBm and −9.6 dBm, respectively, at the HD-FEC threshold. These results highlight the APD’s strong competitiveness for high-bandwidth and energy-efficient data transmission applications.

Wavelength division multiplexing measurements

Finally, to demonstrate the WDM capability and channel uniformity of the proposed device, we employed an eight-channel APD array (Fig. 6a) in combination with an external DWDM demultiplexer with a channel spacing of 200 GHz ( ≈ 1.14 nm in the O-band). In the present proof-of-concept experiment, the demultiplexed channels were characterized sequentially. The demultiplexed wavelengths of λ₁ = 1305.92 nm, λ₂ = 1307.06 nm, λ₃ = 1308.20 nm, λ₄ = 1309.34 nm, λ₅ = 1310.48 nm, λ₆ = 1311.62 nm, λ₇ = 1312.76 nm, and λ₈ = 1313.90 nm were successfully detected. Owing to the narrow wavelength span ( < 10 nm) of the eight DWDM channels, the measured responsivity shows negligible wavelength dependence. Clear eye diagrams were obtained for all eight channels at a data rate of 200 Gbps PAM4, as shown in Fig. 6b, demonstrating the broadband capability and good uniformity of the APD array.

Fig. 6: Wavelength-division multiplexing performance of the Si-Ge APD.
Fig. 6: Wavelength-division multiplexing performance of the Si-Ge APD.The alternative text for this image may have been generated using AI.
Full size image

a Optical microscope image of the eight-channel APD array. b Measured clear and open eye diagrams of 8 × 200 Gbps PAM4 signals.

Discussion

Table 1 provides an overview of state-of-the-art Ge APDs and Si-Ge APDs, which primarily operate in the O-band and C-band wavelength ranges. Earlier studies proposed using the Ge region as the multiplication region and integrating all PN junctions within the Ge of the APD5. However, this approach suffers from excessive light absorption by dopants, resulting in insufficient responsivity. To address this limitation, researchers developed a vertically structured Ge APD with simultaneous Si-Ge doping, effectively reducing doping-induced absorption and enhancing responsivity24. Despite these improvements, Ge APDs still face inherent challenges. When Ge is used as the avalanche region, its high impact ionization coefficient ratio k leads to increased noise, dark current and a limited GBP, thereby constraining the viability of Ge APDs as a mainstream solution5.

Table 1 Summary of the state-of-the-art Ge APDs and Si-Ge APDs
Full size table

In contrast, Si-Ge APDs exhibit a lower effective impact ionization coefficient ratio k. As shown in Table 1, a reduced k value is critical for minimizing the excess noise under high multiplication gain, thereby significantly improving the GBP2,5,13,40. In SACM-structured Si-Ge APDs, the engineered electric field distribution confines the multiplication region primarily to the Si layer, further reducing the effective k value and effectively suppressing noise. Based on the electric field orientation, SACM Si-Ge APDs can be classified into vertical and lateral configurations. The vertical SACM Si-Ge APD has undergone substantial advancements through continuous optimization, achieving a relatively low avalanche voltage, which is beneficial for APD applications6,13,14. However, its GBP remains modest, and its fabrication process is more complex. Additionally, direct contact between Ge and the electrode can degrade device performance. In contrast, the lateral SACM Si-Ge APD features a simpler fabrication process, eliminating Ge-electrode contact and thereby preventing responsivity and bandwidth degradation7,20,21,22,23. Through ongoing research efforts, the GBP of lateral SACM Si-Ge APD has been significantly improved while maintaining a bias voltage below 15 V, further enhancing their potential for practical applications.

Given that most silicon photonic receivers are designed for TE-mode detection42,43,44, this study presents a lateral SACM APD specifically optimized for TE-mode detection, achieving a high gain of 244 and an OE bandwidth of 31 GHz. This was accomplished by optimizing the doping regions on both sides of the Ge layer, effectively reducing the electric field strength within Ge while ensuring carrier transport at saturation velocity. Additionally, we observed a significant increase in gain as the input optical power decreased. Specifically, for input optical powers of −12 dBm, −14 dBm, −16 dBm, −18 dBm, −20 dBm, −22 dBm, and −24 dBm, the corresponding gains were 36, 44, 73, 110, 134, 187, and 244, respectively. Notably, this APD surpasses all previously reported APDs in terms of GBP under any input optical power condition. However, the relatively high bias voltage poses a challenge for integration into computing architectures, as most existing architectures require a driving power supply below 12 V or even lower. Therefore, our next objective is to further refine the doping region’s area and positioning to reduce the APD’s operating voltage. Meanwhile, we are exploring the use of optimized inductive gain peaking to enhance bandwidth, aiming for even higher performance.

In conclusion, we present a high-performance lateral SACM Si-Ge APD that is fully compatible with CMOS technology. Operating at a reverse bias voltage of 12.5 V, the device maintains a low dark current of 8.4 µA. In the O-band, with an input optical power of −16 dBm, the APD achieves a gain of 73 and a GBP of 2628 GHz. As the input optical power decreases to −24 dBm, the gain increases to 244, and the GBP reaches a record-breaking 7564 GHz. Furthermore, we demonstrate the device’s exceptional data reception capabilities, successfully detecting 64 Gbps and 100 Gbps NRZ signals, as well as 128 Gbps and 200 Gbps PAM4 signals, at an input optical power of −16 dBm. Additionally, we showcase an 8 × 200 Gbps communication capacity, establishing a strong foundation for future terahertz data transmission. At the HD-FEC threshold, the measured sensitivities for 100 Gbps NRZ and 112 Gbps PAM4 signals are −17.1 dBm and −9.6 dBm, respectively. With its high GBP, high sensitivity, and compatibility with CMOS technology, this APD is promising for next-generation high-speed and energy-efficient optical communication systems.

Methods

Device fabrication

The APD was fabricated on a 130-nm SOI platform, and a schematic cross-section of the device is shown in Fig. 1b. The Si waveguide was formed by etching the 220-nm-thick top Si layer, with an etch depth of 70 nm as shown in the figure. After waveguide etching, ion implantation was performed to dope the Si waveguide. Subsequently, a 260-nm-thick Ge absorption layer was epitaxially grown on the doped Si waveguide. Tungsten was then used to form ohmic contacts to the heavily doped regions, followed by copper metallization for vias and interconnects. Finally, an Al layer was deposited to form the contact pads.

Eye diagram measurements

The experimental setup for the eye-diagram measurement is shown in Fig. 4b. During the measurements, the frequency response of the test link was calibrated and compensated using a Keysight M8194A arbitrary waveform generator (AWG), resulting in a relatively flat frequency response of the entire electrical-optical-electrical link within a 40 GHz bandwidth and effectively minimizing the influence of external link components. A PRBS-11 data pattern (211-1 bits) generated by the AWG was used for all eye-diagram measurements.

The measurement setup employed a full high-speed Radio Frequency (RF) chain, including RF probes, RF cables, bias-tees, TFLN MZMs, and RF amplifiers, all with bandwidths exceeding 67 GHz. No TIA was integrated after the APD, and no external digital signal processing (DSP) or offline equalization was applied. Only the built-in real-time forward equalization (5 taps) and filtering functions of the sampling oscilloscope were enabled during the measurements. These signal-processing functions effectively compensate for frequency roll-off and mitigate inter-symbol interference, thereby providing effective system bandwidth for the successful reception of 200 Gbps PAM-4 signals and enabling clear eye openings, despite the intrinsic small-signal 3-dB bandwidth of the APD being below 40 GHz.

Multiplication gain

The avalanche gain of the APD was extracted from current–voltage measurements under optical power. The avalanche multiplication gain G(V) was defined as the difference between the photocurrent (Ip) and the dark current (Id) at a given bias voltage, normalized to the corresponding value at the unity gain point V0: \({{\rm{G}}}\left({{\rm{V}}}\right)=({{\rm{Ip}}}({{\rm{V}}})-{{\rm{Id}}}({{\rm{V}}}))/({{\rm{I}}}{{\rm{p}}}({\rm{V}_{0}})-{{\rm{I}}}{{\rm{d}}}({\rm{V}_{0}}))\).