IBM researchers presented a 64-core mixed-signal chip that performs neural-network calculations using phase-change memory. Instead of repeatedly transferring stored weights to a separate arithmetic unit, the chip represents those weights through electrical conductance and carries out matrix operations within memory arrays. The design combines this analog computation with digital functions for processing activations and moving information between cores.
The international team included IBM Research Zurich's Manuel Le Gallo and colleagues working on in-memory computing. Manufactured in a 14-nanometre process, the research chip demonstrated inference on networks including image-recognition and recurrent models with accuracy approaching software implementations. Published in Nature Electronics, the work addressed how analog computing could support more complete neural-network workloads rather than isolated operations. The research processor brought analog memory arrays and their supporting digital circuitry together on one chip.
