In 2025 Chipchain prepared a full-custom technical plan for an NPU, the neural processing unit at the heart of an AI chip. The plan is Chipchain's own, and the process node of the plan is N3P, a 3 nm-class leading-edge process.

Chipchain prepared the plan for its work with a company that leads the world in robotics and artificial intelligence.

The plan covers two depths of customisation, MAC array customisation, evaluation against a baseline, block-level sign-off and compute-core test, in four stages of work from technical evaluation to full-custom design.

An NPU spends most of its area and power in its MAC arrays, the grids of multiply-accumulate units that do a neural network's arithmetic. The more copies of a block a chip carries, the more each saving in that block is worth, so the arrays are where full-custom design pays first.

Evaluation against a baseline means building the same block twice, once with standard cells and automatic place and route, and once with the full-custom techniques, then comparing area, power and speed under the same conditions. The comparison comes first, so each later stage is decided on evidence.

The techniques are the company's full-custom methods: custom cells, library characterisation and re-characterisation, hand-crafted netlists, manual place and route and custom clock trees.