CMOS scaling has been the backbone of the overall logic roadmap for decades, but it is reaching its physical limits while also imposing ever more constraining design restrictions. This has triggered a critical need for new device architectures and integration concepts to be able to continue delivering profitable node-to-node scaling gains and to help preserve the industry’s power-performance-area-cost metrics. From the transistor’s perspective, vertically stacked lateral nanosheet (NS) FETs, with a gate-all-around (GAA) configuration, are widely regarded as the most promising and mature option to replace finFETs. Reduced gate lengths should be feasible thanks to their improved electrostatics, thus allowing further scaling of the contacted-gate-pitch and of the cell height via a reduced number of metal tracks. Other key characteristics include high design flexibility, with various NS widths possible on a given wafer, and larger drivability per layout footprint by increasing the number of vertically stacked NS per device and/or using wider NS [1,2] (Fig.1). An extension of this technology could in principle be envisioned by strongly reducing the p-n separation in the so-called forksheet configuration [3]. Beyond that, the concept of stacking devices with different polarity on top of each other is also being looked at [4,5]. Other future technology candidates include FETs with vertical transport [6] and non-silicon channels [7]. Each new architecture will have its own specific challenges such as the internal routeability for stacked structures in functional logic blocks (e.g., standard cell or SRAM) but, in general, many elements can be shared by the various branches of the NS family of devices.

Overall, a careful balance between drive strength and capacitance is required in NS FETs engineering. In particular, the presence of dielectric inner spacers in-between vertically stacked nanosheets is a critical element, also as it leads to a different growth regime for the source/drain (S/D) epi as compared to the situation in finFETs [8]. This is an important differentiator as channel strain induced by S/D has been traditionally used to boost device performance. The feasibility of continuing using process-induced stress techniques for mobility enhancement is in fact a key challenge for several new architectures, namely for the top device in stacked structures or when S/D are placed in different vertical levels. Moreover, faced with power scaling stagnation, cold computing is also becoming an attractive option to consider for enabling high performance boosting in an energy efficient way. Our results confirm improved DC properties for NS FETs (e.g., subthreshold swing (SS), mobility), with similar mechanisms responsible for their noise behavior at room and low temperatures (300K (RT), 78K) [9].

In addition to the need for the introduction of new transistor technologies, given the increased complexity and cost in back-end-of-line processing, it has also become ever more pressing to address both wiring and power delivery network (PDN) bottlenecks to take full advantage of the scaling performance benefits at transistor level. The concept of moving the PDN to the wafer’s backside (BS) such that it can alleviate routing congestion on its frontside (FS) has been recently gaining traction [10,11]. This is illustrated in Fig.2 wherein, by combining logic and 3D technologies, both wafer sides are used. In our work, after frontside processing, device and carrier wafers are bonded at RT, including a 523K post-bond anneal. Extreme wafer thinning is then implemented prior to nano-through-silicon-vias (n-TSV) definition (landing on the metal-1 level (M1) in the frontside) and backside metallization. Evaluating the impact on scaled transistors from BS processing, our results show similar p/n threshold voltages (V_Ts) can be obtained with an extra sinter at the end of fabrication. Inclusion of an additional high-pressure H₂-anneal prior to the final sinter is also seen to help lower the SS values for pmos without significant I_OFF effect. Reliability-wise, constant ramped voltage stress measurements also show no BTI degradation for p/nmos, with additional indication of potential benefits by the final anneal(s) treatment selection. These findings are further corroborated by LF-noise analysis.

References

[1] N. Loubet et al., VLSI Tech. Dig., 2017, p.230.

[2] A. Veloso et al., SSDM Tech. Dig., 2019, p.559.

[3] P. Weckx et al., IEDM Tech. Dig., 2019, p.871.

[4] W. Rachmady et al., IEDM Tech. Dig., 2019, p.697.

[5] C.-Y. Huang et al., IEDM Tech. Dig., 2020, p.425.

[6] A. Veloso et al., IEDM Tech. Dig., 2019, p.230.

[7] P.-C. Shen et al., Nature, 2021, Vol.593, p.211.

[8] G. Eneman et al., ECS Trans., 2020, Vol.98(5), p.253.

[9] B. Cretu et al., EuroSOI-ULIS Tech. Dig., 2021.

[10] A. Veloso et al., VLSI Tech. Dig., 2021, TFS2-6.

[11] https://www.intel.com/content/www/us/en/events/accelerated.html.