Why WDM for inference is not just a networking transplant
Wavelength division multiplexing has been a cornerstone of fiber-optic communications for decades. The idea is straightforward: different data streams are modulated onto different optical wavelengths, then combined onto a single fiber and separated at the receiver. The fiber carries several times the data of a single-wavelength link at the same propagation speed.
When applied to on-chip photonic inference, the motivation looks superficially similar but the engineering constraints diverge significantly. In a WDM telecom link, the wavelengths are well-separated, carefully stabilized, and the modulation rates are high because bandwidth is the scarce resource. In a photonic inference chip, what we want is not more bandwidth on a data link. We want to run multiple independent matrix-vector products simultaneously on the same physical waveguide mesh, using wavelength as the index. The resource being multiplied is compute, not communication capacity.
The distinction matters because it changes which parameters you optimize. A telecom WDM system is built to maximize bits per second with low bit-error rate. A WDM inference channel is built to maintain computation accuracy across all wavelength channels simultaneously, with a very different tolerance budget for crosstalk.
Encoding parallel inference streams on separate channels
The way we think about WDM for inference is as follows. A photonic MZI matrix is, at its core, a linear optical transformation. If you launch light at wavelength lambda_1 into the chip, the chip performs one matrix-vector product on the optical field at that wavelength and outputs the result at the detectors. If you simultaneously launch light at lambda_2, and the chip's MZI settings have the same transfer function at lambda_2 as at lambda_1, then the chip is performing two identical operations in parallel, one for each stream.
More practically: the two wavelength channels can carry different input vectors while experiencing the same weight matrix encoded in the MZI phase settings. This means a single chip simultaneously processes multiple inference requests that share a model. For a data-center inference server running many concurrent requests against the same model weights, this is exactly the parallelism that matters. Throughput scales with channel count without loading new weights or reconfiguring the mesh.
This is not saying you can run different models simultaneously. The weight matrix is set by the phase-shifter configuration, and all wavelength channels see the same phase settings. WDM parallelism is about request throughput for a fixed model, not multitasking across different models. That distinction is worth stating clearly to avoid overstating what the architecture delivers.
Wavelength spacing and crosstalk in practice
The practical limit on how many wavelength channels you can use is set by crosstalk. In a silicon waveguide, two wavelengths propagating simultaneously interact through several mechanisms. The most significant in the inference context is the wavelength dependence of the MZI transfer function itself. An MZI's splitting ratio is a function of the directional coupler geometry and the operating wavelength. If the coupling coefficient varies by even a few percent across the wavelength channels, different channels see slightly different matrix elements, and the accuracy of all channels degrades together.
In silicon photonics, the coupler dispersion in the C-band (around 1550 nm) is gentle enough that a few channels separated by 10 to 20 nm can see similar coupling coefficients. Wider separation or more channels require dispersion-engineered couplers, which add design complexity. Our current work operates in a regime where we can maintain acceptable per-channel accuracy with a small number of channels, rather than maximizing channel count.
There is also inter-channel optical crosstalk from imperfect wavelength-selective components such as multiplexers and demultiplexers. On-chip silicon photonics ring resonator add-drop filters have a finite isolation, typically in the range of 20 to 30 dB. The residual leakage of one channel into another appears as a systematic offset error in the output of the adjacent channel. Managing this requires either careful channel spacing, filter design with higher isolation, or digital calibration that estimates and subtracts the crosstalk contribution at the detector.
The demultiplexer at the output
One part of the architecture that is easy to underestimate is what happens after the waveguide mesh. In a single-wavelength system, the photodetectors simply read the output ports of the MZI network. In a WDM system, each output port carries all wavelength channels superimposed. Before the photodetectors can read each channel independently, the wavelengths must be separated.
On-chip demultiplexing using arrayed waveguide gratings (AWGs) or ring resonator filter banks adds insertion loss. An AWG with 4-channel demultiplexing in silicon might contribute an additional 2 to 4 dB of excess loss per channel, depending on fabrication quality and the number of channels. This loss comes directly out of the signal-to-noise ratio at the detector and reduces the effective dynamic range of the computation.
For this reason, the number of useful WDM channels is not just a function of laser availability or chip real estate. It is bounded by how much additional loss the system can absorb before the detector's photocurrent becomes too small to distinguish from noise. Every channel added requires a noise floor analysis, not just a layout area calculation.
Thermal implications of multi-wavelength operation
A point that often goes undiscussed in WDM inference proposals is thermal management. Running multiple wavelengths simultaneously means the laser sources contribute more aggregate optical power to the chip. In a regime where two-photon absorption in silicon is negligible (which holds at the power levels typical of inference chips), the added thermal load is small relative to the electronic power dissipation. But the MZI thermal phase shifters are temperature-sensitive. If the chip's thermal map shifts due to changing optical input power, the MZI calibration drifts.
We have seen this in measurements: a chip calibrated at steady state with a single channel shows phase drift when a second channel is added, because the added optical power raises the waveguide temperature slightly and shifts the effective refractive index. The magnitude is small, but it is non-zero, and the calibration firmware needs to account for it. This is another instance of the broader principle that the device, the system, and the runtime are not separable engineering problems in photonic inference hardware.
Where we are with WDM
We are actively characterizing multi-wavelength operation in our current test chips, working with 2 to 4 channel configurations. The results are at the device-characterization stage; we are not reporting throughput numbers because any numbers we could give now would reflect a research setup with manual alignment, not a packaged system. What we can say is that the physics is tractable, the main engineering work is in the demux loss budget and the calibration firmware, and both are understood problems rather than open research questions. WDM will be part of the GSK-1 architecture. The channel count and per-channel accuracy specs will come when we have them in a form worth stating.