🔬 optics

Nonlinear pluggable optics: Digital signal processing-free Intensity Modulated Direct Detection links using analog photonic Next Generation Reservoir Computing

This paper proposes a Nonlinear Pluggable Optic (NLPO) transceiver utilizing analog photonic Next-Generation Reservoir Computing on a photonic integrated circuit to compensate for channel impairments, demonstrating through simulations that it outperforms traditional DSP-based solutions in range, latency, and power efficiency for 50 GBd PAM-4 IM/DD links.

Original authors: Nicholas Cox, Joseph Murray, Ross T. Schermer, Shuo S. Pang, Christopher Long, Gajadhar Joshi, Nicholas Nobile, Raktim Sarma, Brandon Redding

Published 2026-06-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicholas Cox, Joseph Murray, Ross T. Schermer, Shuo S. Pang, Christopher Long, Gajadhar Joshi, Nicholas Nobile, Raktim Sarma, Brandon Redding

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a complex message (like a high-speed video) through a long, twisty tunnel (an optical fiber). In the world of data centers, the current standard is to use a "digital signal processor" (DSP) at the receiving end. Think of the DSP as a very smart, but slow and power-hungry, translator who has to sit at the end of the tunnel, listen to the garbled message, and use a massive amount of brainpower to figure out what the original words were.

The problem is that as data speeds get faster, this translator gets overwhelmed. The message gets distorted by the tunnel itself (due to physics like dispersion and nonlinearity), and the translator needs more power and time to fix it. This creates a bottleneck for the AI and machine learning clusters that need instant data.

The New Solution: The "Nonlinear Pluggable Optic" (NLPO)

The authors of this paper propose a new device called a Nonlinear Pluggable Optic (NLPO). Instead of using a slow, power-hungry digital translator, they use a "photonic reservoir computer."

Here is how it works, using some everyday analogies:

1. The "Magic Mirror" vs. The "Smart Translator"

The Old Way (DSP): Imagine the message arrives at the end of the tunnel as a blurry, scrambled photo. The digital translator has to look at the photo, run complex math algorithms, and try to reconstruct the original image. It's accurate but takes time and electricity.
The New Way (Photonic NGRC): The authors built a device that acts like a magic mirror. When the distorted light hits this mirror, the mirror doesn't just reflect it; it naturally "unscrambles" the pattern instantly as the light bounces off. It doesn't need to calculate anything; the physics of the mirror itself does the work. This happens at the speed of light, with almost no power.

2. Seeing the Invisible (The Phase Trick)

One of the biggest hurdles in sending data is that standard receivers only see the "brightness" (intensity) of the light, not its "phase" (the timing and shape of the wave). It's like trying to understand a song by only looking at the volume knob, ignoring the melody and rhythm.

The Paper's Claim: The new device can "see" the hidden phase information without needing a bulky, expensive reference laser (called a local oscillator).
The Analogy: Imagine trying to identify a person in a crowd. The old way is like only seeing their shadow (intensity). The new way allows the device to see the person's face (phase) directly, even though it's just looking at the shadow. This extra information makes it much easier to fix the scrambled message.

3. The "Super-Speed Camera" (Oversampling)

To fix a distorted message, you usually need to take many samples of it.

The Problem: In a digital system, taking samples faster than the data is sent is like trying to take a photo of a hummingbird's wings with a camera that only takes one picture per second. You miss everything. To do this digitally, you need incredibly fast (and expensive) electronics.
The Paper's Claim: The photonic device can "oversample" the data naturally.
The Analogy: Instead of a camera that takes one photo per second, the new device is like a strobe light that flashes thousands of times in the blink of an eye. It captures the entire motion of the data wave in high definition without needing expensive, high-speed electronic cameras.

What Did They Actually Prove?

The authors ran computer simulations to test this idea. They didn't build the physical device yet, but they modeled it perfectly. Here is what their simulation showed:

Distance: They tested a system sending data at 50 GBd (a very fast speed).
- Old LPO (Linear Pluggable Optic): Could only send data about 1 km before the signal became too garbled to fix.
- Old DSP (Digital): Could reach about 3 km.
- New Photonic NGRC: Successfully sent data over 50 km while keeping the error rate low enough for real-world use.
Performance: The new system didn't just match the digital one; it actually outperformed the best digital methods available, even when the digital system was given more "brainpower" (memory) to try to fix the errors.
Realism: They added "noise" (like static on a radio) and "loss" (signal fading) to the simulation to mimic real-world conditions. Even with these challenges, the system worked as long as they added a standard amplifier (EDFA) to boost the signal.

Summary

The paper proposes a way to send high-speed data much further and faster without using the heavy, power-hungry digital computers we currently rely on. By using a specialized chip that manipulates light directly (analog photonic reservoir computing), they can "unscramble" distorted messages instantly, seeing hidden details that digital systems miss, and doing it all with a fraction of the energy.

The authors conclude that this approach could enable data centers to connect over tens of kilometers with low latency and low power, a feat currently impossible with standard linear optics or difficult to achieve with digital processing.

Technical Summary: Nonlinear Pluggable Optics via Analog Photonic Next-Generation Reservoir Computing

Problem Statement
The escalating bandwidth demands of AI/ML data center interconnects have driven the adoption of Linear Pluggable Optics (LPO) to reduce power consumption and latency by eliminating Digital Signal Processing (DSP). However, LPO architectures lack the ability to correct nonlinear distortions arising from non-ideal encoding electronics (DACs, modulators) and optical channel impairments (chromatic dispersion, Kerr nonlinearity). Consequently, LPOs are restricted to the O-band (near-zero dispersion) and shorter ranges. Conversely, DSP-based solutions, while capable of longer reach, consume significant power (up to 50% of transceiver power) and introduce unacceptable latency for real-time applications. Furthermore, standard Intensity Modulation/Direct Detection (IM/DD) links lose optical phase information, limiting DSP's ability to compensate for distortions beyond ~1 km at high symbol rates (e.g., 50 GBd).

Methodology
The authors propose a Nonlinear Pluggable Optic (NLPO) transceiver that utilizes an analog photonic Next-Generation Reservoir Computing (NGRC) architecture implemented on a Photonic Integrated Circuit (PIC). Unlike traditional reservoir computing, NGRCs do not rely on recurrent feedback; instead, they generate feature vectors through linear and nonlinear transformations of discrete time samples.

The proposed PIC-based NGRC operates as follows:

Input Feature Vector Generation: The incoming distorted optical field is split into $N_{in}$ paths with varying delays to create a time-delayed input vector. Crucially, this vector preserves the complex optical field (amplitude and phase) without requiring a local oscillator (LO).
Feature Mixing: The input vector is coherently mixed in a passive region (e.g., evanescently coupled waveguides) to form a speckle pattern. This mixing effectively decomposes the signal into polynomial combinations of the input data, serving as the reservoir state.
Weighted Detection: Trained weights are applied to the mixed optical power using micro-ring modulators. The weighted signals are detected by balanced photodetectors, and the resulting photocurrent is subtracted to recover the corrected symbols.
Oversampling: The system oversamples the data at rates ( $R_{os}$ ) significantly higher than the symbol rate (e.g., 250 GS/s for a 50 GBd link) using optical delay lines, bypassing the need for high-bandwidth electronic sampling.

Key Contributions

DSP-Free Nonlinear Compensation: The paper demonstrates a fully analog, DSP-free approach to compensating for both electronic encoding distortions and optical propagation impairments (dispersion and Kerr nonlinearity).
Phase Information Utilization without LO: The work shows that the photonic NGRC can access and utilize optical phase information for equalization without the need for a local oscillator or co-propagating carrier, a capability typically reserved for coherent systems.
Optical Oversampling: The authors identify that the ability to optically oversample data well above the symbol rate is a major advantage of photonic reservoirs, enabling the extraction of information that digital DSP cannot access without prohibitively expensive electronics.
PIC Implementation: The study details a spatially multiplexed PIC architecture suitable for integration with existing transceiver technologies, emphasizing passive feature vector generation to minimize power consumption.

Results
Simulations were conducted on a 50 GBd PAM-4 link (100 Gb/s) to evaluate performance against DSP-based FIR filters and digital NGRCs:

Performance at 10 km: In a noise-free 10 km link, the photonic NGRC significantly outperformed DSP-based methods. While digital FIR filters and digital NGRCs failed to meet the KP4 Forward Error Correction (FEC) threshold ( $2.2 \times 10^{-4}$ ), the photonic NGRC achieved a Bit Error Rate (BER) below this threshold with a memory length ( $M$ ) of 6 or greater.
Impact of Oversampling: Increasing the oversampling rate ( $R_{os}$ ) allowed the photonic NGRC to achieve error-free transmission ( $BER < 10^{-6}$ ) with $R_{os} \ge 2$ (with LO) or $R_{os} \ge 3$ (without LO). DSP-based approaches failed to reach the FEC threshold regardless of oversampling rate.
Range Extension: By optimizing memory length, the photonic NGRC extended the viable link range from the typical 1 km limit of LPOs to over 50 km. Specifically, with $M=11$ , the system maintained a BER below the KP4 threshold up to 51.2 km.
Optimization for Real-World Constraints: The authors optimized the design for a practical PIC by reducing the output feature vector size ( $N_{out}$ ) to 32. An optimal configuration of $M \approx 4.2$ and $R_{os} = 5$ achieved a BER of $2 \times 10^{-5}$ .
Noise Robustness: Simulations including Amplified Spontaneous Emission (ASE) and shot noise demonstrated that with an Erbium-Doped Fiber Amplifier (EDFA) providing ~20 dB gain, the system could overcome PIC insertion losses and maintain BER below the FEC threshold.

Significance and Claims
The paper claims that this work showcases the potential for photonic NGRCs to outperform state-of-the-art digital solutions in real-world applications. The proposed NLPO architecture offers a path to low-latency, low-power IM/DD links capable of operating over tens of kilometers (10s of km) and across the C-band, overcoming the range and robustness limitations of current LPOs. The authors emphasize that this is the first work to demonstrate that oversampling is a primary advantage of photonic reservoir computers. While the current results are simulation-based and include noise modeling, the work establishes a theoretical and architectural foundation for experimental realization, which the authors note is their current focus.