The physics of light-based inference
This page covers how Mach-Zehnder interferometer arrays perform matrix multiplication in coherent light, why optical dot products consume less energy per operation than transistor switching, and how Great Sky structures these components into a working inference co-processor.
From transistor gates to interference fringes
Every multiply-accumulate in a transformer layer ultimately moves charge through a transistor. Charge movement involves resistive heating, capacitive loading, and clock-cycle delays. Those costs scale with model size and inference throughput in a way that is difficult to engineer around.
Mach-Zehnder interferometers as multipliers
A Mach-Zehnder interferometer splits an optical input into two paths, applies a phase shift to one arm via an electro-optic modulator, then recombines them. The amplitude of the output is a function of the phase difference between arms. By setting that phase shift to represent a matrix weight, the device performs a continuous multiplication in the optical domain without switching any transistors in the data path.
In a mesh of cascaded MZI units, each MZI represents one element in a unitary transform. Shen et al. (2017) demonstrated that any unitary matrix can be decomposed into a product of two-port MZI operations, which means an arbitrary matrix multiplication can be expressed as a physical light-propagation problem through a structured waveguide mesh.
Coherent light and linear algebra
The key physical property is that coherent light obeys linear superposition. When two optical fields meet at a beam splitter or directional coupler, their amplitudes add vectorially. The result at each output port is a linear combination of the inputs, with coefficients determined by the splitting ratio and phase. This is, by definition, a dot product.
A neural-network weight vector encoded as optical amplitude and phase values, propagated through a waveguide mesh, produces at the output an exact matrix-vector product. The computation happens during propagation at the speed of light. There is no clock edge, no memory fetch for weights during the computation, and no resistive switching in the matrix-multiply path.
Encoding weights and inputs
Weights are set once per inference context by programming the electro-optic modulators in the MZI mesh to specific phase states. Input activation vectors are encoded as optical amplitude values injected into the waveguide inputs. The mesh computes the product as the light propagates. Photodetectors at the output ports read the result and hand it back to the digital host as a current proportional to optical power.
Four functional blocks on one substrate
The Great Sky photonic processor integrates all four stages of optical inference on a single silicon photonic die, connected via PCIe to the host compute node.
High-speed electro-optic modulators set the phase and amplitude of coherent light to encode weight values. Each modulator covers one weight slot in the mesh.
A structured MZI mesh performs the matrix-vector product. Light propagates through the mesh passively; no switching occurs during the multiply-accumulate phase.
Photodetectors at each output waveguide convert optical intensity to photocurrent. Each detector output represents one element of the result vector.
Integrated analog-to-digital converters digitize the photodetector outputs and produce a result vector ready for the host's inference framework via PCIe DMA.
The physical reason optical multiply-accumulate is more efficient
The energy cost of a digital multiply-accumulate operation comes from three sources: fetching operands from memory, performing the arithmetic in transistor logic, and dissipating the heat those operations produce. Optical matrix multiplication removes one of the three entirely and reduces a second.
Transistors dissipate energy every time they switch state. In a digital MAC unit, the weight must be loaded from memory, routed through arithmetic logic, and the transistors in that path switch with every operation. Photons propagating through a phase-set waveguide mesh perform the same multiplication with no switching in the compute path.
The modulator that sets a weight phase draws power only when the weight is updated, not during inference. Once the phase is set, the optical field propagates through the mesh passively. There is no capacitive load to charge and discharge on each operation, no resistive heating in a switching element, and no current through a gate dielectric.
The energy cost that remains is in the electro-optic driver circuitry, the photodetectors, and the ADC. These are bounded by the signal-to-noise requirements of the output precision, not by the dimensionality of the matrix being multiplied. Larger matrices in the same waveguide mesh do not proportionally increase the energy-per-operation cost the way they do in digital systolic arrays.
Built on a decade of photonic computing research
The field of photonic neural networks and optical matrix accelerators has been developing since the mid-2010s in academic and national laboratory settings. Great Sky is an engineering effort to take findings from that research base and apply them to the practical constraints of data-center inference hardware. The references below represent the published foundations of this approach.
Great Sky's engineering work addresses the practical gap between these results and production deployment: thermal stability, fiber coupling loss at PCIe integration, and the software interface that lets existing inference frameworks use optical compute without code changes.