Glossary

Each definition is only as broad as the linked guide. The same abbreviation can mean different devices in other industries.

Analog-to-digital converter (ADC): An electronic circuit that measures an analog signal and represents it as digital numbers. In an analog photonic accelerator it can digitise detector outputs; its speed, precision, and power are part of the system cost. Discussed in The practical limitations of optical computing.
Co-packaged optics (CPO): Packaging optical engines close to an electronic switch or processor within the same package. The shorter electrical connection can help data movement. CPO describes how a connection is built, rather than a claim that the processor performs calculations optically. Discussed in Co-packaged optics, optical I/O, and pluggable transceivers.
Coprocessor: A processor that handles selected tasks for a host CPU. A photonic coprocessor may accelerate particular mathematical operations while the host manages the program, data, and operations the device cannot perform. Discussed in What is an optical processing unit?.
Digital-to-analog converter (DAC): An electronic circuit that turns digital numbers into an analog voltage or current. Some photonic systems use DACs to drive modulators or other controls; not every optical computing architecture uses the same conversion path. Discussed in How photonic computing performs a calculation.
Graphics processing unit (GPU): An electronic processor designed to perform many operations in parallel. GPUs support graphics and many numerical workloads, including machine learning; useful comparisons depend on the workload, precision, software, and complete system being measured. Discussed in Photonic accelerators versus GPUs: workloads and trade-offs.
Inference: Running a trained model to produce an output for new input, such as classifying an image. Training adjusts the model's parameters; accelerating one inference operation does not by itself accelerate the entire training or inference pipeline. Discussed in Photonic accelerators versus GPUs: workloads and trade-offs.
Matrix-vector multiplication (MVM): Multiplying each row of a matrix by a vector and summing the products to produce an output vector. It appears in many numerical and machine-learning workloads and is an operation some optical architectures implement. Discussed in How photonic computing performs a calculation.
Native Processing Unit (Q.ANT NPU): Q.ANT's name for its photonic processor, which performs mathematical operations using optical signals. Here NPU means Native Processing Unit; the same abbreviation elsewhere can mean a neural processing unit with a different architecture. Discussed in What is an optical processing unit?.
Numerical precision: How finely a system can represent or distinguish numerical values. An analog device's effective precision depends on noise, calibration, and measurement conditions; a claimed bit count should be checked against the accuracy needed by the workload. Discussed in How to evaluate photonic-computing performance claims.
Optical input/output (optical I/O): Interfaces that use optical signals to transfer data into and out of a chip, package, or computing system. Optical I/O can connect electronic processors while their arithmetic remains electronic. Discussed in Co-packaged optics, optical I/O, and pluggable transceivers.
Optical interconnect: A connection that carries data using light between components or systems. Its purpose is communication, such as linking switches or computing devices; carrying the input to a calculation is distinct from performing that calculation. Discussed in Optical processors versus optical interconnects.
Optical processing unit (OPU): A descriptive name for hardware that performs selected calculations using light. The term does not specify one universal instruction set or architecture; LightOn has used OPU for an accelerator based on optical random-feature transformations. Discussed in What is an optical processing unit?.
Photonic computing: Computing in which optical effects carry out mathematical operations. A practical system may combine an optical calculation with electronic memory, control, conversion, and other processing; the exact division depends on the architecture. Discussed in How photonic computing performs a calculation.
Photonic integrated circuit (PIC): A chip containing connected optical components such as waveguides, modulators, and detectors. A PIC can support communication or computation, and may use materials including silicon or lithium niobate. Discussed in Optical processors versus optical interconnects.
Pluggable optical transceiver: A removable module that converts electrical data to optical signals for transmission and converts received optical signals back to electrical data. It connects through a compatible equipment port and optical interface. Discussed in Co-packaged optics, optical I/O, and pluggable transceivers.
Random projection: Mapping an input vector using randomly chosen coefficients, often to build features or a smaller mathematical sketch. An optical device may implement a fixed random transform; this is different from supporting arbitrary user-selected matrix coefficients. Discussed in What is an optical processing unit?.
Silicon photonics: Optical components built with silicon-based fabrication processes, often integrated with electronic circuits. It is a technology platform for data links and some computing devices; photonic processors can also use other material platforms. Discussed in Where photonics fits in an AI data centre.
System benchmark: A measurement of a defined task across the complete path needed to produce a useful result. For a photonic accelerator this can include host work, data transfer, conversion, calibration, and the optical operation, with accuracy and power boundaries stated. Discussed in How to evaluate photonic-computing performance claims.