Story details

A preprint presents SIXPPAC, a mixed-signal machine-learning engine for real-time integrated sensing and communication in wireless baseband. Its design combines two successive-approximation ADCs with a six-layer digital binary-weight neural network.

The authors report a 0.765-square-millimetre 12-nanometre FinFET implementation operating at 105 million samples per second and 4.18 nanojoules per inference. These are paper-reported prototype measurements, not a commercial product specification.