IB Computer Science Topic A4.1.2: Hardware Requirements for Various Scenarios Where Machine Learning Is Deployed Key Definitions

An application-specific integrated circuit (ASIC) is a chip designed and fabricated to carry out one fixed task. It cannot be reprogrammed after manufacture, which is what makes it fast and power-efficient on that task and useless if the task changes.

Edge computing processes data on or beside the device that produced it, rather than sending it to a central data centre first. Removing that round trip is what allows a response in milliseconds.

A field-programmable gate array (FPGA) is a chip whose internal logic can be reconfigured after manufacture by loading a new design onto it. One board can be rewired for a new model, at the cost of being slower and less power-efficient than an ASIC doing the same job.

Inference is the stage where a trained model is given a new input and returns a prediction. Each input passes through the model once, so inference needs far less processing and storage than training the same model did.

The ability of a system to handle more data and more users by adding further nodes rather than by replacing existing hardware with something more powerful.

A tensor processing unit (TPU) is a processor built for the tensor arithmetic that neural network layers are made of. It beats a general-purpose GPU on that specific work, and is usually rented through a cloud platform rather than owned.