The Future of CMOS: More Moore or a New Disruptive Technology?

Abstract

For more than four decades, Moore’s law has been driving the semiconductor industry where the number of transistors per chip roughly doubles every 18–24 months at a constant cost. Transistors have been relentlessly evolving from the first Ge transistor invented at Bell Labs in 1947 to planar Si metal-oxide semiconductor field-effect transistor (MOSFET), then to strained SiGe source/drain (S/D) in the 90and 65-nm technology nodes and high-κ/metal gate stack introduced at the 45and 32-nm nodes, then to the current 3D transistors (Fin field-effect transistors (FinFETs)) introduced at the 22-nm node in 2011 (Figure 1.1). In extremely scaled transistors, the parasitic and contact resistances greatly deteriorate the drive current and degrade the circuit speed. Thus, miniaturization of devices so far has been possible due to changes in dielectric, S/D, and contacts materials/processes, and innovations in lithography processes, in addition to changes in the device architecture [1, 2]. The gate length of current transistors has been scaled down to 14 nm and below, with over 109 transistors in state-of-the-art microprocessors. Yet, the clock speed is limited to 3–4 GHz due to thermal constraints, and further scaling down the device dimensions is becoming extremely difficult due to lithography challenges. In addition, further scaling down the complementary metal-oxide semiconductor (CMOS) technology is leading to larger interconnect delay and higher power density [3]. The complexity of physical design is also increasing with higher density of devices. So, what is next? A promising More-than-Moore technology is the 3D integrated circuits (ICs) which can improve the performance and reduce the intra-core wire length, and thereby enable high transfer bandwidth with reduced latencies and power consumption, while maintaining compact packing densities [4]. Alternative technologies that could be promising for new hardware accelerators include resistive computing, neuromorphic computing, and quantum computing. Resistive computing could lead to non–von Neumann (VN) computing and enforce reconfigurable and data-centric paradigms due to its massive parallelism and low power consumption [5]. Moreover, humans can easily outperform current high-performance computers in tasks like auditory and pattern recognition