A bit-parallel molybdenum disulfide computer built through multi-level co-optimization

Abstract

The sustainable development of artificial intelligence requires energy-efficient computing technology. Two-dimensional semiconductor digital electronics could potentially provide such capabilities, and promising single devices and small circuits have been developed. However, very large-scale integration remains challenging due to the inability to control atomic-scale defects and mesoscopic device variations, as well as the lack of macroscopic variation-aware design methodology. Here we report a molybdenum disulfide computer that combines a 0.5-μm industrial fabrication process and a back-end-of-line-integrated academia laboratory process. The computer comprises 1,433 transistors interconnected by four metal layers within a compact footprint, offering an integration density of around 9,336 transistors per square millimetre. The computer can store data on-chip in the register file and perform arithmetic operations on multiple-bit data parallelly at a 1-kHz clock frequency. Key to the development of this system is a multi-level co-optimization methodology that spans transistor, standard cell, logic synthesis and interconnect design.