Due to their growing popularity and computational cost, deep neural networks (DNNs) are being targeted for hardware acceleration. A popular architecture for DNN acceleration, adopted by the Google Tensor Processing Unit (TPU), utilizes a systolic array based matrix multiplication unit at its core. This paper deals with the design of faulttolerant, systolic array based DNN accelerators for high defect rate technologies. To this end, we empirically show that the classification accuracy of a baseline TPU drops significantly even at extremely low fault rates (as low as 0.006%). We then propose two novel strategies, fault-aware pruning (FAP) and fault-aware pruning+retraining (FAP+T), that enable the TPU to operate at fault rates of up to 50%, with negligible drop in classification accuracy (as low as 0.1%) and no run-time performance overhead. The FAP+T does introduce a one-time retraining penalty per TPU chip before it is deployed, but we propose optimizations that reduce this one-time penalty to under 12 minutes. The penalty is then amortized over the entire lifetime of the TPU's operation.

Reverse engineering (RE) is one of the major security threats to the semiconductor industry due to the involvement of untrustworthy parties in an increasingly globalized chip manufacturing supply chain. RE efforts have already been successful in extracting device level functionalities from an integrated circuit (IC) with very limited resources. Camouflaging is an obfuscation method that can thwart such RE. Existing work on IC camouflaging primarily involves transformable interconnects and/or covert gates where variation in doping and dummy contacts hide the circuit structure or build cells that look alike but have different functionalities. Emerging solutions, such as polymorphic gates based on a giant spin Hall effect and Si nanowire field effect transistors (FETs), are also promising but add significant area overhead and are successfully decamouflaged by the satisfiability solver (SAT)-based RE techniques. Here, we harness the properties of two-dimensional (2D) transition-metal dichalcogenides (TMDs) including MoS2, MoSe2, MoTe2, WS2, and WSe2 and their optically transparent transition-metal oxides (TMOs) to demonstrate area efficient camouflaging solutions that are resilient to SAT attack and automatic test pattern generation attacks. We show that resistors with resistance values differing by 5 orders of magnitude, diodes with variable turn-on voltages and reverse saturation currents, and FETs with adjustable conduction type, threshold voltages, and switching characteristics can be optically camouflaged to look exactly similar by engineering TMO/TMD heterostructures, allowing hardware obfuscation of both digital and analog circuits. Since this 2D heterostructure devices family is intrinsically camouflaged, NAND/NOR/AND/OR gates in the circuit can be obfuscated with significantly less area overhead, allowing 100% logic obfuscation compared to only 5% for complementary metal oxide semiconductor (CMOS)-based camouflaging. Finally, we demonstrate that the largest benchmarking circuit from ISCAS’85, comprised of more than 4000 logic gates when obfuscated with the CMOS-based technique, is successfully decamouflaged by SAT attack in <40 min; whereas, it renders to be invulnerable even in more than 10 h when camouflaged with 2D heterostructure devices, thereby corroborating our hypothesis of high resilience against RE. Our approach of connecting material properties to innovative devices to secure circuits can be considered as a one of a kind demonstration, highlighting the benefits of cross-layer optimization.

Hyperpolarized singlet states provide the opportunity for polarization storage over periods significantly longer than T1. Here, we show how the singlet state in a chemically equivalent proton spin system can be revealed by a weak power spin-lock. This procedure allowed the measurement of the lifetimes of the singlet state in protic solvents. The contributions of different intra- and intermolecular relaxation mechanisms to singlet lifetimes are investigated with this procedure.

No abstract