Summary
Today’s neural networks run on digital systems that consume significant power, limiting the deployment of advanced AI in size-, weight-, and power- (SWaP-) constrained environments. Analog in-memory computing promises greater energy and area efficiency, but current approaches are often hampered by power-hungry analog-to-digital converters and environmental circuit sensitivities.
The Scalable Analog Neural-networks (ScAN) program is addressing these challenges by designing analog neural networks that connect directly to analog sensor outputs, thereby eliminating the need for data converters and achieving orders-of-magnitude improvements in energy efficiency. Launched in 2025 as a 54-month, two-phase e fort, ScAN will first demonstrate robust, intermediate-scale systems before scaling to large networks.