Alex Yoon, Intermolecular business of Merck KGaA, Darmstadt, Germany
Technologies such as gate all around FETs and neuromorphic analog memory will increase speed and reduce power consumption in the chiplets that control artificial intelligence of things devices and systems.
While 5G provides improved speed to consumers, it’s the industrial, automotive, infrastructure, and IoT applications that make 5G stand out. These applications will need IoT devices that will rely on 5G and AI, thus creating an artificial intelligence of things (AIoT).
The world needs billions of new smart systems built with ICs that connect devices through wired systems and wireless systems. Wireless power and signal transfers from edge AIoT systems opens the greatest potential for global transformation. AIoT devices and systems have, however, some of the most challenging specifications for low-power operation and reliability. Reaching targets of less than 1 mW operating power, with less than 1 ms latency, in a chip smaller than 1 mm in size will require the integration of new semiconductor materials.
To achieve the AIoT goal, we’ll need faster RF power amplifiers, specialized AI/ML server clouds, and a network edge consisting of sensors, actuators, and microcontrollers. IC chips or “chiplets” for each of these functions will be heterogeneously integrated into complete systems for specific applications. While many semiconductor processes such as GaN can overtake silicon, we’ll need even more new materials in the coming years.
The 5G New Radio standard includes both the sub-6 GHz band and the 28 GHz to 41 GHz bands. The mmWave frequencies enable not just faster data rates but step-function increases in bandwidth, latency, and reliability needed for AIoT.
5G hardware foundations on different chips
RFICs capable of sending and receiving data at 5G speeds are usually made from specialized “III-V” (or “3-5” referring to columns in the periodic table of the elements). The III-V compound semiconductor materials include gallium-nitride (GaN) and gallium-arsenide (GaAs) on 150 mm wafers. Fabricated on 200 mm wafers, silicon and silicon-carbide (SiC) RFICs provide relatively reduced functionalities at lower cost. Silicon processes will continue to support some 5G applications.
RF power amplifiers made from GaN and GaAs have many performance advantages over traditional Si-based semiconductors, such as higher switching speed, lower electrical current loss resulting from a low RDS(ON), and higher power density. Figure 1 shows a cross-section of a III-V High Electron Mobility Transistor (HEMT) integrated with silicon CMOS by the Singapore MIT Alliance for Research & Technology (SMART) using wafer-scale layer transfer and bonding. GlobalFoundries recently licensed Raytheon’s gallium-nitride on silicon (GaN-on-Si) technology for 5G and far-future 6G commercial wireless applications, which will be available on 200 mm GaN-on-Si wafers from GF’s Fab 9 in Essex Junction, Vermont.
Logic chips for server racks “in the cloud” will evolve to increase bandwidth for different workloads, with an increasing number of cores using CMOS technology. Commercial CMOS logic includes planar FETs, finFETs, and soon gate all around (GAA) channels branded as RibbonFET, horizontal nanosheet (HNS), or Horizontal Nanowire (HNW).
Memory for 5G cloud centers will evolve using digital volatile and nonvolatile memory (NVM) technologies, optimized for different applications with storage-class memory and compute-in-memory (CIM) materials and devices. Memory for the 5G edge will generally need NVM to save power, and neuromorphic analog memory blocks could reduce power consumption by 1000x compared to the best digital NVM.
Chiplets allow for the easiest heterogeneous integration (HI) of diverse 5G technologies within a single package, with each chiplet made on a different high-volume manufacturing (HVM) fab line:
- GaN and GaAs for RF and opto-electronics,
- Silicon CMOS on 300 mm wafers for high-speed Digital Logic and AI/ML,
- Silicon on 200 mm wafers for low-power Analog AI/ML,
- NVM including CIM, SCM, and Neuromorphics,
- Sensors and Actuators specialized for each application, and
- Antennas and Passive devices.
Figure 2 shows a simplified schematic of an edge AIoT system that could add reliability and security to infrastructure such as buildings, electrical networks, bridges, and tunnels. Vibration sensors with extremely low power consumption and localized AI can detect signals of impending failures. Such detection can avoid unnecessary preventative maintenance while alerting when service or replacement is needed.