The shift to smaller nodes in the near future is going to be driven by the rise of 5G technology. The beginning of 5G equipped devices such as the Samsung Galaxy S10 5G, Samsung Galaxy Note 10+ 5G, OnePlus 7 Pro 5G and Huawei Mate 20 X 5G has coincided with widespread adoption of EUV technology. This has equipped foundries with the tools and more importantly the demand for chipsets utilising smaller nodes as consumers are keen to adopt new communication technology. Consequently, TSMC plans to roll out its 5 nm chipsets a year earlier than previously expected by the IRDS roadmap. FinFET is the most suited geometry at 5 nm since it is cost effective and a much more mature technology when compared to Si-nanowire. Additionally, at 5 nm, parasitic properties are still reasonably low.
Demand for smaller semiconductor nodes is expected to continue through 2022, where FinFET will no longer be able to support a stable transistor design as the fin’s ability to have good enough electrostatic control over the channel is greatly reduced. This is due to the drain voltage creating SCEs such as DIBL as the channel length is shortened. Therefore, this may be the first year, as the node is shrunk to 3 or 4 nm, that we see lateral gate-all-around nanowire (LGAA) designs as the only viable alternative factoring in cost, maturity and manufacturing difficulty. LGAA is expected to mitigate FinFET’s biggest problem at this node; this is the lack of control over the channel. By wrapping the gate all around the channel, electrostatic control will be greatly improved, leading to a much more stable transistor design.
Approaching 2024 and 2025, scaling will start to become exceedingly difficult. We will see efforts to work around this problem through two solutions. The first involves a switch to vertical gate-all-around nanowire (VGAA) which should be relatively inexpensive since foundries will likely gain sufficient experience with nanowire technology with the manufacture of LGAA for two years. VGAA works around the problem of scaling, through the separation of gate length scaling from foot-print scaling. Thus, the overall footprint will be able to shrink to perhaps two more nodes without terrible SCEs and parasitics. According to the IRDS guidelines, 2024 may see the use of III-V material, however I disagree since experimental data is still far from what has been predicted in theoretical models. This is due to problems regarding yield and quality of III-V materials during manufacture.
The second method to work around scaling in 2024 will be the use of 3D architecture, more specifically, 3D TSV based chipsets. This technology is very likely since gate length scaling alone will likely not deliver the performance enhancements expected by consumers from node-to-node. The introduction of 3D TSV will likely be the most important technology to rescue Moore’s Law in the near future since we already have experience with the technology in the shape of SSDs and manufacturing techniques have plenty of time to be ready and perfected by 2024 for logic devices. There is simply no cost effective alternative to 3D TSV in the coming years.
The time period leading up to 2030 will likely be relatively calm and uninteresting. As 2030 approaches, the scaling potential of Si-CMOS will reach its absolute limit, even with VGAA technology. Thus, we will start to see the implementation of beyond CMOS technology, where silicon is replaced. By this time, it is expected that nanowire technology has been perfected with silicon and III-V material may finally be ready. Thus, the two will probably be combined to form III-V compound semiconductors. This will primarily increase charge carrier flow, leading to higher operating frequencies, increasing computational power.
Shortly after 2030, the potential of 3D TSV will have been most likely reached, as more layers simply cannot be stacked on top due to spatial and thermal limitations as a result of the thickness of the layers and low TSV density. By this time, monolithic 3D technology should be a cost effective alternative to 3D TSV, since bottom-up manufacturing techniques should be mature as GAA technology is widespread. Monolithic 3D will greatly increase the number of layers present, massively increasing transistor density.
During this period, although high performance computing may not see dramatic change, low performance-high efficiency computing will see two new technologies: NEMs and TFET/SE-TFET. The node size for the year 2030 is expected to be around 1.5-2 nm. At this length, I believe that NEMs simply cannot match the raw switching speed and therefore performance of TFET. However NEMs may still be an incredibly important technology since it will be one of the only computational technologies to withstand environmental conditions such as ionising radiation. With NASA planning to send humans to Mars in the year 2030 and private companies such as SpaceX building permanent residences on Mars around the same timeframe, NEMs maybe key to building computer systems capable of surviving the harshest of conditions.
The quantity of data collected is growing exponentially and is thus predicted to reach an estimated 175 zettabytes by just 2025. TFETs will be useful in the storage, retrieval and management of the huge quantities of data which will be collected due to the rise of IoT and technology giants as well as our governments continuing to collect vast sums of information about our daily activities. TFETs are the only viable low power solution due to their ability to utilise the otherwise detrimental quantum effects such as quantum tunneling which happen as gate length approaches 1 nm.
From the period between 2035 and 2040, the need for sub-nanometer technologies may become apparent. For high performance solutions, there may be no better alternative than CNT FETs. Due to the superb quantum confinement effects and carrier transport properties exhibited by CNTs, they may be the only way to avoid worsening SCEs and extend performance further through scaling. By this time, CNTs should be easily manufactured in large quantities with great yields; their manufacturing difficulty is one of the greatest hurdles holding them back today. Other parts of the transistor may also be replaced by graphene, TMDCs or other 2D materials. Beyond 2040, classical computing as a whole may no longer be able to cope with the computational needs of the world.
The exponential increase in computing power can be continued through the extension of CMOS and development of
Beyond CMOS technologies for many decades more. Long live Moore's Law!