Autonomous learns Connect-4 ASIC on 130nm
The author explores the concept of automating the design of custom silicon chips for arbitrary machine learning tasks, arguing that extreme hardware specialization could soon emerge if production costs fall. As a proof of concept, the author presents a single-cycle, autonomously learned Connect-4 chip comprising about 14,000 logic gates. Built on a 130 nm manufacturing process, the chip runs at 25 MHz and achieves energy efficiency several orders of magnitude greater than a CPU-based engine. Open article →