Cognition and higher order reasoning in the human brain have been shown to adhere closely to probabilistic inference frameworks such as Bayesian networks that support reasoning under uncertainty. We architect a physically equivalent Bayesian network fabric with nanotechnology, employing inherently stochastic spintronic devices in unique recursive analog circuit structures that support Bayesian inference through physical fabric properties. This fabric approach results in many orders of magnitude efficiency improvements over conventional approaches and enables new cognitive applications with millions of random variables that are not possible today.
Our European Union’s Horizon-2020 project aims to develop a complete synthesis and performance optimization
methodology for switching nano-crossbar arrays that leads to the design and construction of an emerging nanocomputer.
Within the project, we investigate different computing models based on either two-terminal switches, realized with field effect
Nano-crossbar arrays have emerged as area and power efficient structures with an aim of achieving high performance computing beyond the limits of current CMOS. Due to the stochastic nature of nano-fabrication, nano arrays show different properties both in structural and physical device levels compared to conventional
Hardware implementations of artificial neural networks (ANNs) have become feasible due to the advent of persistent 2-terminal devices such as memristor, phase change memory, MTJs, etc. Hybrid memristor crossbar/CMOS systems have been studied extensively and demonstrated experimentally. In these circuits, memristors located at each cross point in a crossbar are, however, stacked on top of CMOS circuits using back end of line processing (BOEL), limiting scaling.
Non-volatile 3D FPGA research to date utilizes layer-by-layer stacking of 2D CMOS / RRAM circuits. On the other hand, vertically-composed 3D FPGA that integrates CMOS and RRAM circuits has eluded us, owing to the difficult requirement of highly customized regional doping and material insertion in 3D to build and route complementary p- and n-type transistors as well as resistive switches. In the layer-by-layer nonvolatile 3D FPGA, the connectivity between the monolithically stacked RRAMs and underlying CMOS circuits is likely to be limited and lead to large parasitic RCs.
Monolithic 3D IC (M3D) shows degradation in performance compared to 2D IC due to the restricted thermal budget during fabrication of sequential device layers. A transistor-level (TR-L) partitioning design is used in M3D to mitigate this degradation. Silicon validated 14nm FinFET data and models are used in a device-to-system evaluation to compare the TR-L partitioned M3D’s (TR-L M3D) performance against the conventional gate-level (G-L) partitioned M3D’s performance as well as standard 2D IC.
Design for power-delivery network (PDN) is one of the major challenges in 3D IC technology. In the typical layer-by-layer stacked monolithic 3D (M3D) approaches, PDN has limited accessibility to the device layer away from power/ground source due to limited routability and routing resources in the vertical direction. This results in an incomplete and low-density PDN design and also severe IR-drop issue. Some improved M3D approaches try to enlarge design area to create additional vertical routing resources for robust and high-density PDN design.
Gene Expression Networks (GENs) attempt to model how genetic information stored in the DNA (Genotype) results in the synthesis of proteins, and consequently, the physical traits of an organism (Phenotype). Deciphering GENs plays an important role in a wide range of applications from genetic studies of the origins of life to personalized healthcare. Probabilistic graphical models such as Bayesian Networks (BNs) are used to perform learning and inference of GENs from genetic data.
Parallel and monolithic 3D integration directions realize 3D integrated circuits (ICs) by utilizing layer-by-layer implementations, with each functional layer being composed in 2D. In contrast, vertically-composed 3D CMOS has eluded us likely due to the seemingly insurmountable requirement of highly customized complex routing and regional 3D doping to form and connect CMOS pull-up and pull-down networks in 3D. In the current layer-by-layer directions, routing can be worse than 2D CMOS because of the limited pin access.