What is an optical processing unit?

An optical processing unit, or OPU, uses light to perform a computational operation. For this guide, the term means a photonic accelerator that transforms input data into a calculated result. Understanding that operation matters more than the label on the chip.

Start with the operation

Consider a processor that repeatedly calculates weighted sums: multiply several input numbers by corresponding weights, then add the products. A photonic implementation represents those inputs in optical signals and arranges an optical system so its measured output corresponds to the sum.

This is a different job from carrying a message over fibre. A fibre link aims to reproduce transmitted data at the receiver. A computing device aims to produce a specified transformation of that data.

Published examples include programmable nanophotonic circuits for neural-network operations and a photonic tensor core for convolution processing. They illustrate different designs rather than one universal OPU architecture. Shen et al., Feldmann et al.

When reading an announcement, finish this sentence: “The optical component takes ___ as input and calculates ___.” If the answer only describes bandwidth between chips, the announcement concerns an interconnect.

A small calculation an OPU might accelerate

Suppose a simple scoring system has three inputs and two possible outputs. Use these invented values:

input x = [2, 1, 3]

weights for output A = [ 0.5, 1.0,  0.0]
weights for output B = [-0.5, 0.0,  1.0]

A = 0.5×2 + 1.0×1 + 0.0×3 = 2
B = -0.5×2 + 0.0×1 + 1.0×3 = 2

Together these are a matrix-vector multiplication, often written y = Wx. The matrix W contains the two rows of weights; x is the input vector.

The arithmetic example is exact on paper. A physical accelerator must also decide how to encode positive and negative values, scale the signals, and recover outputs within an acceptable error. Optical intensity alone cannot represent a negative number without an additional encoding scheme.

The useful feature is the possibility of implementing many contributions to a weighted sum together. It does not imply that every instruction in a program can be converted into the same optical circuit.

How numbers become light

Different designs use different signal properties. These may include optical power, field amplitude and phase, multiple spatial paths, or separate wavelengths. A device changes those signals through configured optical components; detectors then provide measurements.

The mechanism is architecture-specific. Shen and colleagues use coherent nanophotonic processing, while Feldmann and colleagues combine phase-change memory arrays with an optical frequency comb. Hamerly and colleagues describe photoelectric multiplication using coherent detection. These papers are useful entry points for seeing why “photonic computing” covers several physical approaches. Shen et al., Feldmann et al., Hamerly et al.

A useful reading question is where the weights live. They might be stored in an optical material state, represented by tunable device settings, or supplied through another signal. The answer affects how quickly a new workload can be loaded.

Where the rest of the computer goes

A practical system still needs a way to accept a job, obtain its data, configure the accelerator, and return results. For a hybrid design, the flow can look like this:

application
    ↓
host prepares data and schedules supported operations
    ↓
input conversion → photonic operation → output measurement
    ↓
digital post-processing and application result

This is a conceptual diagram, not a specification for every product. Some systems take optical inputs directly; others use electronic conversion at their boundaries.

For a cloud user, the visible interface might be a software library or a job endpoint. That interface should explain supported operators, shapes, precision, and error handling. Calling a familiar software function does not establish that the job ran on physical photonic hardware: a simulator may expose a similar interface.

Three claims to keep separate

This table is an editorial framework for reading evidence. A research demonstration should be judged by its experiment; a service should additionally be judged by its operating contract.

What would make an OPU useful?

The question is whether an entire task finishes with acceptable quality, time, and energy. That includes preparing data, moving it, configuring hardware, and processing measurements.

McMahon’s analysis explains why the speed of light alone is an inadequate argument for a computing advantage. Architecture and system design determine whether optical properties become useful benefits. The physics of optical computing

Start with one supported operation and one real workload. Then ask for results at equal output quality, including the surrounding system. That approach makes an OPU easier to understand and turns a broad technology promise into a question that can be tested.