Paper Analysis | A Symmetrical, Wavelength Agnostic, Bidirectional, Silicon-Photonic Link Proposal and Demonstration | NVIDIA

2 days ago

4 min read

In the era of exploding AI compute, GPU clusters are growing exponentially, putting enormous pressure on switch port count (radix). Co-packaged optics (CPO) has drawn attention for saving significant power and cost, but the physical "beach-front" at the edge of a CPO substrate is limited, severely capping how far port count can scale.

How do you double throughput without adding beach-front? Bidirectional transmission (BiDi) is a strong answer: it doubles the system's port count and halves fiber connection cost per bit. But traditional BiDi solutions are either too expensive or too cumbersome. Today we break down a new NVIDIA paper that uses a 65nm silicon photonics process to propose a "hardware-symmetrical, wavelength-agnostic" polarization-multiplexed BiDi architecture. In short, they built a technology that lets the equipment at both ends be identical and communicate bidirectionally at high speed without having to assign "wavelength colors"!

Paper Details

Title: A Symmetrical, Wavelength Agnostic, Bidirectional, Silicon-Photonic Link Proposal and Demonstration
Authors: SHAI COHEN, LIRON GANTZ, SEGEV ZARKOVSKY, RONI BAR, HANANEL FAIG, BENJAMIN G. LEE AND C. THOMAS GRAY
Affiliation: NVIDIA Corporation

Figure-by-Figure Analysis: Inside NVIDIA's Technology

The heart of this paper is its architecture design and validation method, so let's walk through the figures:

Figure 1: Three BiDi Architectures Head to Head

This figure highlights the industry's current pain points and why NVIDIA proposed a new architecture:

(A) Traditional circulator-based architecture: It keeps the equipment at both ends (Alice and Bob) identical, but circulators significantly increase packaging complexity and component cost.
(B) Wavelength-multiplexed (color-based) architecture: This is the common approach today, using a silicon-photonics-integrated wavelength multiplexer (MUX) to assign different wavelengths to the upstream and downstream links. But it causes serious practical problems: the two ends must be "color-coded" (e.g., one end transmits red and receives green, the other transmits green and receives red). That forces network designers to plan every connection in advance and even stock multiple switch variants, making inventory management a nightmare.
(C) Polarization-multiplexed (Polarization MUX) architecture: This is the paper's killer move. By separating upstream and downstream by polarization state, both ends can be completely identical (no color distinction). Since in any standard single-mode fiber (SMF) BiDi scheme both ends must already be able to receive both polarization states, why not use that directly for bidirectional transmission!

Figure 2: The Math and Architecture Behind Polarization Tracking

Here NVIDIA explains why the equipment at both ends can be fully symmetrical:

Theoretical basis: The polarization evolution of light propagating through single-mode fiber can be described by an SU(2) Jones transformation matrix.
Decoding mechanism: Once the polarization tracker at the receiver (Bob) fully compensates the channel's relative phase difference and rotates the state, the channel's overall transfer matrix becomes strictly anti-diagonal.
Perfect bidirectional mapping: Matrix algebra shows that the conjugate transpose of this anti-diagonal matrix is still anti-diagonal. That means 100% of Alice's signal lands on Bob's receive port, and conversely 100% of Bob's signal returns to Alice's receive port, achieving a perfect cross-connection.
Symmetrical hardware design: Figure 2.B shows that by placing a polarization tracker at both ends (one actively tracking, the other idle), the hardware can be fully symmetrical, putting an end to the annoying "wavelength pairing" problem.

Figure 3: Experimental Setup

To prove this isn't just theory, the team built a validation platform:

Hardware configuration: Alice and Bob use the same physical architecture, connected to single-mode fiber (SMF) through 2D grating couplers (2D GC).
Test conditions: Each end independently injects a 32 Gbps NRZ (PRBS15) signal at 1299nm.
Disturbance emulation: A polarization scrambler is inserted in the middle to emulate the dynamic disturbances fiber experiences in the real world. Alice continuously adjusts thermo-optic phase shifters (TOPS) through a controller to maximize received power, while Bob stays idle.

Figure 4: Emulating a Harsh Polarization-Disturbance Environment

To prove the tracker is fast enough, Figure 4 shows the "torture" metrics applied to the channel:

(A) Disturbance spectrum: The measured disturbance power spectral density (PSD) shows band-limited "white noise" characteristics with a 3dB frequency of 42Hz.
(B) Angular-velocity statistics: The probability density function (PDF) of polarization angular velocity shows a mean of 94.5 rad/s and a 95th percentile of 161.4 rad/s. That means the tracker must stay rock-solid under extremely fast polarization changes.

Figure 5: 32 Gbps Test Results

The experimental results are excellent:

(A) & (B) Bathtub curves: Both directions (Tx1->Rx2 and Tx2->Rx1), transmitting simultaneously at 32 Gbps, achieve a bit error rate (BER) below 1e-11. The small difference between the two curves is purely due to modulators from different vendors at each end.
(C) & (D) Eye diagrams: The eyes in both directions are clean and wide open, further confirming distortion-free transmission.
Bonus highlight: The team also switched Bob's laser to 1299.5nm with no performance degradation, proving the architecture is truly wavelength agnostic. TOPS power is below 50mW, and each tracker's insertion loss is only about 0.25dB.

Conclusion: A Colorless Future for AI Compute Networks

This NVIDIA work neatly demonstrates how polarization multiplexing on a 65nm silicon photonics process can unlock bidirectional transmission over single-mode fiber. By decoupling the control mechanism from the data rate (tuning relies only on received power), the technique can be seamlessly extended to more complex modulation formats and different laser grids in the future.

For future AI data centers, this means doubling switch port count and halving fiber cost without adding to the burden on the switch package edge. More importantly, "identical hardware at both ends" will fully solve the headaches network planners face in equipment deployment and inventory management. This is a very promising path for future CPO and high-density optical communications!