Photonic Computing Theory
Mathematical foundations of optical neural network accelerators
Contents
1. Optical Computing Fundamentals
Photonic computing leverages the unique properties of light for computation: high bandwidth through wavelength-division multiplexing (WDM), low-latency signal propagation, and energy-efficient matrix operations via optical interference.
1.1 Why Photonics?
The key advantages of photonic accelerators over electronic counterparts:
- Parallelism: WDM enables N operations simultaneously on N wavelengths
- Speed: Light propagates at ~200 μm/ps in silicon waveguides
- Energy: Passive optical interference requires no switching energy
- Bandwidth: THz modulation frequencies are achievable
1.2 Key Physical Principles
Electromagnetic Wave in Waveguide:
$$E(z,t) = E_0 e^{i(\beta z - \omega t)}$$
where β = neffk₀ is the propagation constant, neff is the effective refractive index, and k₀ = 2π/λ is the free-space wavenumber.
Optical Power Detection:
$$P = |E|^2 = E \cdot E^*$$
The square-law detection of photodetectors is what enables the multiplication operation: when two signals are combined and detected, the output contains their product.
2. Ring Resonator Theory
Microring resonators are the fundamental building blocks for photonic weights. They act as wavelength-selective filters whose transmission can be tuned to implement different weight values.
2.1 Transfer Function
Through-port Transmission:
$$T = \frac{a^2 - 2at\cos\phi + t^2}{1 - 2at\cos\phi + (at)^2}$$
where:
- $a = e^{-\alpha L/2}$ is the round-trip amplitude transmission
- $t = \sqrt{1-\kappa^2}$ is the self-coupling coefficient
- $\phi = \beta L = 2\pi n_{eff} L / \lambda$ is the round-trip phase
- $\kappa$ is the power coupling coefficient
2.2 Key Parameters
Free Spectral Range (FSR):
$$FSR = \frac{\lambda^2}{n_g L}$$
Quality Factor:
$$Q = \frac{\lambda}{\Delta\lambda_{FWHM}} = \frac{\pi n_g L \sqrt{at}}{\lambda(1-at)}$$
Finesse:
$$\mathcal{F} = \frac{FSR}{\Delta\lambda_{FWHM}} = \frac{\pi\sqrt{at}}{1-at}$$
2.3 Weight Implementation
By tuning the resonance wavelength (via thermal or electro-optic effects), we control the transmission at the signal wavelength. This transmission directly implements the weight value:
Weight Encoding:
$$w = \sqrt{T(\lambda_{signal})} \in [0, 1]$$
For signed weights, differential encoding uses two resonators: $w = w^+ - w^-$
3. Matrix-Vector Multiplication
Photonic MVM is the core operation for neural network inference. It exploits wavelength parallelism and optical summation to perform N² multiply-accumulate operations in O(1) time.
3.1 Crossbar Architecture
Matrix-Vector Product:
$$y_i = \sum_{j=1}^{N} W_{ij} x_j$$
In a photonic crossbar:
- Inputs $x_j$ are encoded on wavelengths $\lambda_j$ (amplitude modulation)
- Weights $W_{ij}$ are implemented by ring resonator transmission at (row i, wavelength j)
- Summation occurs via optical combining and balanced photodetection
3.2 Signal Flow
Optical Field at Output i:
$$E_{out,i} = \sum_{j=1}^{N} \sqrt{P_j} \cdot x_j \cdot \sqrt{T_{ij}} \cdot e^{i\phi_{ij}}$$
Detected Photocurrent:
$$I_i = R_{pd} \cdot |E_{out,i}|^2 = R_{pd} \sum_j P_j x_j^2 T_{ij}$$
3.3 Compute Throughput
Operations per Second:
$$TOPS = N^2 \cdot 2 \cdot f_{clock} \times 10^{-12}$$
For an N×N crossbar operating at frequency fclock, we get 2N² operations per cycle (N² multiplies + N² accumulates).
4. Phase-Change Materials
Phase-change materials (PCMs) like Ge₂Sb₂Te₅ (GST) provide non-volatile optical memory for storing neural network weights without continuous power consumption.
4.1 Optical Properties
Complex Refractive Index:
$$\tilde{n} = n + ik$$
The dramatic change in optical constants between amorphous and crystalline states enables large modulation:
| Material | namor | ncryst | Δn | kcryst |
|---|---|---|---|---|
| GST | 4.0 | 6.5 | 2.5 | 1.5 |
| GSST | 3.2 | 4.8 | 1.6 | 0.2 |
4.2 Effective Medium Theory
Bruggeman Approximation:
$$c \frac{\epsilon_{cryst} - \epsilon_{eff}}{\epsilon_{cryst} + 2\epsilon_{eff}} + (1-c) \frac{\epsilon_{amor} - \epsilon_{eff}}{\epsilon_{amor} + 2\epsilon_{eff}} = 0$$
where c ∈ [0,1] is the crystalline fraction, enabling multi-level weight storage.
4.3 Programming Energy
SET (Crystallization) Energy:
$$E_{SET} = P \cdot t_{pulse} \approx 10\text{ mW} \times 100\text{ ns} = 1\text{ nJ}$$
5. Noise Analysis
Understanding noise sources is critical for determining the achievable precision in photonic computing systems.
5.1 Shot Noise
Shot Noise Current (RMS):
$$i_{shot} = \sqrt{2qI_{ph}B}$$
where q is electron charge, Iph is photocurrent, and B is bandwidth.
5.2 Thermal Noise
Johnson-Nyquist Noise:
$$i_{thermal} = \sqrt{\frac{4k_BT \cdot B}{R_L}}$$
5.3 Relative Intensity Noise (RIN)
RIN Noise:
$$i_{RIN} = I_{ph}\sqrt{RIN \cdot B}$$
5.4 Total Noise & SNR
Total Noise:
$$i_{total} = \sqrt{i_{shot}^2 + i_{thermal}^2 + i_{RIN}^2}$$
Signal-to-Noise Ratio:
$$SNR = 20\log_{10}\left(\frac{I_{signal}}{i_{total}}\right) \text{ dB}$$
Effective Number of Bits:
$$ENOB = \frac{SNR - 1.76}{6.02}$$
6. Thermal Effects
Silicon's large thermo-optic coefficient makes thermal management critical for photonic computing systems.
6.1 Thermo-Optic Effect
Refractive Index Change:
$$\Delta n = \frac{dn}{dT} \Delta T \approx 1.86 \times 10^{-4} \cdot \Delta T \text{ (for Si)}$$
6.2 Resonance Shift
Wavelength Shift:
$$\Delta\lambda = \frac{\lambda}{n_g} \frac{dn}{dT} \Delta T \approx 80 \text{ pm/K}$$
This means a 1°C temperature change shifts the resonance by ~80 pm, which can significantly affect weight accuracy.
6.3 Thermal Stabilization Power
Heater Power for Tuning:
$$P_{heater} = \frac{\Delta\lambda}{80 \text{ pm/K}} \times P_{\pi}$$
where Pπ ≈ 10-50 mW is the power required for a full FSR shift, depending on heater efficiency.
6.4 Thermal Time Constant
Response Time:
$$\tau_{th} = R_{th} \cdot C_{th} \approx 1-100 \text{ μs}$$
The thermal time constant limits how fast weights can be reconfigured, making PCM-based non-volatile weights attractive for inference.
References
- Shen, Y., et al. "Deep learning with coherent nanophotonic circuits." Nature Photonics 11, 441-446 (2017)
- Feldmann, J., et al. "All-optical spiking neurosynaptic networks." Nature 569, 208-214 (2019)
- Wuttig, M., et al. "Phase-change materials for non-volatile photonic applications." Nature Photonics 11, 465-476 (2017)
- Bogaerts, W., et al. "Programmable photonic circuits." Nature 586, 207-216 (2020)
- Miscuglio, M., et al. "Photonic tensor cores for machine learning." Applied Physics Reviews 7, 031404 (2020)