Liquid Silicon: A Nonvolatile Fully Programmable Processing-In-Memory Processor with Monolithically Integrated ReRAM for Big Data/Machine Learning Applications

Abstract

A nonvolatile fully programmable processing-in-memory (PIM) processor named Liquid Silicon (L-Si) is demonstrated, which combines the superior programmability of general-purpose computing devices (e.g. FPGA) and the high power efficiency of do-main-specific accelerators. Besides the general computing applications, L-Si is particularly well suited for AI/machine learning and big data applications, which not only pose high computational/memory demand but also evolves rapidly. L-Si is fabricated by monolithically integrating HfO2 resistive RAM on top of commercial 130nm Si CMOS. Our measurement confirmed the fabricated chip operates reliably at low voltage of 650 mV. It achieves 60.9 TOPS/W in performing neural network inferences and 480 GOPS/W in performing content-based similarity search (a key big data application) at nominal voltage supply of 1.2V, showing >$3\times $ and ∼$100\times $ power efficiency improvement over the state-of-the-art domain-specific CMOS-/RRAM-based accelerators. In addition, it outperforms the latest nonvolatile FPGA in energy efficiency by ∼$3\times $ in general compute-intensive applications.