With shrinking device geometries, the device-to-device (D2D) variability of emerging memory devices, affect the performance of deep neural networks (DNN) and restrict the use of standalone devices to be used as Analog in-Memory Computing (AiMC) accelerators. For FeFET based AiMC designs, such effects can be controlled by using resistors or capacitors connected in series with FeFETs. However, in both cases, the effect of variability will again start dominating when the programming window shrinks further due to scaling towards smaller nanometer dimensions. Additionally, while resistance based solutions will add to the existing IR drop issues within the macro, capacitor based approach avoids such a case. In this paper, we evaluate a FeFET and capacitor based AiMC accelerator and study how finetuning cell operating parameters along with design modifications of the AiMC macro, can help in achieving a variability aware AiMC design. Using a 22-nm CMOS PDK augmented with FeFET model based on measured data, evaluation on a DNN with CIFAR-10 benchmark, demonstrated that following our approach, the implemented DNN achieves an accuracy of ∼85% which is an improvement of ∼13% over an untuned AiMC macro.

Analog in Memory Computing (AiMC) based neural network acceleration is a promising solution to increase the energy efficiency of deep neural networks deployment. However, the quantization requirements of these analog systems are not compatible with state-of-the-art neural network quantization techniques. Indeed, while the quantization of the weights and activations is considered by modern deep neural network quantization techniques, AiMC accelerators also impose the quantization of each Matrix Vector Multiplication (MVM) result. In most demonstrated AiMC implementations, the quantization range of MVM results is considered a fixed parameter of the accelerator. This work demonstrates that dynamic control over this quantization range is possible but also desirable for analog neural networks acceleration. An AiMC compatible quantization flow coupled with a hardware aware quantization range driving technique is introduced to fully exploit these dynamic ranges. Using CIFAR-10 and ImageNet as benchmarks, the proposed solution results in networks that are both more accurate and more robust to the inherent vulnerability of analog circuits than fixed quantization range based approaches.

Extremely energy-efficient convolutional neural network inference (CNN) is recently enabled by analog in-memory compute (AiMC). The integration of AiMC in a primarily digital inference system brings new challenges ranging from device specifications to defining novel system architectures. A novel framework to evaluate the impact of AiMC array at the system level is presented. The framework is used to model a SRAM-based 1024 × 512 prototype AiMC array, capable of energy efficiency of upto 675 TMACs/W. The proposed framework allows modelling different compute cells, array dimensions, operating voltage, and activation buffer energy models which can be used to determine the overall energy efficiency for various CNN workloads.

Designers typically add design margins to semiconductor memories to compensate for aging. However, the aging impact increases with technology downscaling, leading to the need for higher margins. This results into a negative impact on area, yield, performance, and power consumption. As an alternative, mitigation schemes can be developed to reduce such impact. This paper proposes a mitigation scheme for the memory's sense amplifier (SA); the scheme is based on creating a skew in the relative strengths of the SA's cross-coupled inverters during design. The skew is compensated by aging due to unbalanced workloads. As a result, the impact of aging on the SA is reduced. To validate the mitigation scheme, the degradation of the sense amplifier is analyzed for several workloads. The experimental results show that the proposed mitigation scheme reduces the degradation of the sense amplifier's critical figure-of-merit, the offset voltage, with up to 26%.

Memory designs typically contain design margins to compensate for aging. As aging impact becomes more severe with technology scaling, it is crucial to accurately predict such impact to prevent overestimation or underestimation of the margins. This paper proposes a methodology to accurately and efficiently analyze the impact of aging on the memory's digital logic (e.g., timing circuit and address decoder) while considering realistic workloads extracted from applications. To demonstrate the superiority of the methodology, we analyzed the degradation of the L1 data and instruction caches for an ARM v8-a processor using both our methodology as well as the state-of-the-art methods. The results show that the existing methods may significantly over-or underestimate the impact (e.g., the decoder margin up to 221% and the access time up to 20%) as compared with the proposed scheme. In addition, the results show that in general the instruction cache has the highest degradation. For example, its access time degrades up to 9% and its decoder margin up to 44%.

Designers typically add design margins to memories to compensate for their aging. As the aging impact increases with technology scaling, bigger margins become necessary. However, this negatively impacts area, yield, performance, and power consumption. Alternatively, mitigation schemes can be used to reduce the impact of aging. This paper proposes a hardware-based mitigation scheme for the memory's address decoder logic. The scheme is based on adapting the decoder's workload during idle cycles by stressing the short paths and putting long paths into relaxation. Thanks to the adapted workload, the impact of aging on the address decoder is reduced, resulting in a more reliable memory. To validate the benefit of the mitigation scheme, the decoder's degradation of the L1 data and instruction caches for an ARM v8-a processor is analyzed. The experimental results show that the proposed mitigation scheme reduces the degradation of the decoder's timing margin with up to 4.1x at negligible area and no more than 3% power overhead.

Designers typically add design margins to compensate for chip aging. However, this leads to yield loss (in case of overestimation) or low reliability (in case of underestimation). This paper analyzes the impact of aging on a complete high-performance industrial 14-nm FinFET SRAM. It investigates the impact on the memory's parametric (i.e., its delay) and functional (i.e., correct functionality) metrics. Moreover, it examines which components are the main contributors to the degradation of the memory's reliability and how it is impacted by workload and environmental conditions, i.e., temperature and voltage fluctuations. This paper not only investigates the impact of the memory's components individually, which is typically the case in prior work, but it also studies the contribution of components' interaction to the overall memory aging. The results show that the timing circuit, address decoder, and the output latches and buffers are the main contributors to the memory's parametric degradation, while the cell, sense amplifier, and address decoder are the main contributors to its functional degradation. Moreover, the results show that it is crucial to consider the impact of the interaction of components on the aging; individual analysis leads to overly pessimistic results and even wrong conclusions in certain cases.

Memory designs usually add design margins to compensate for chip aging; this may lead to yield and performance loss (in case of overestimation) or reduced reliability (in case of underestimation). This paper analyzes the impact of aging on cutting edge high performance 14nm FinFET SRAM using a calibrated aging model; it does not only analyze the impact of the SRAM's components individually, as it is the case in prior work, but it also investigates the contribution of the interaction of these components while considering different workloads; both the overall metric of the memory (i.e., the access time) as well as metrics of individual components (e.g., sensing delay for the sense amplifier) are examined. The results show that it is crucial to consider not only the aging of all individual components, but also their interaction in order to provide accurate prediction of aging effects; considering only aging of single/individual components leads to either too optimistic or pessimistic results. For example, using our approach (which includes the components interaction) results approximately in 9.1% degradation of memory access time (for three years of aging), while using the traditional approach (based on adding the impact of individual components) results in 7.3% increase only; a relative difference of 25%, for which the timing and the address decoder components are the main contributors. With respect to individual components, the sense amplifier is the most fragile one (e.g., its offset voltage spec. degrades up to 58%).

To compensate for time-zero (due to process variation) and time-dependent (due to e.g. Bias Temperature Instability (BTI)) variability, designers usually add design margins. Due to technology scaling, these variabilities become worse, leading to the need for bigger design margins. Typically, only worst-case scenarios are considered, which will not present the actual workload of the targeted application. Alternatively, mitigation schemes can be used to counteract the variability. This paper presents a run-time design-for-reliability scheme for memory Sense Amplifiers (SAs); SAs are an integral part of any memory system and are very critical for high performance. The proposed scheme mitigates the impact of time-dependent variability due to aging by using an on-line control circuit to create a balanced workload. The simulation results show that the proposed scheme can reduce the most critical figures-of-merit, namely the offset voltage shift and the sensing delay of the SA with up to ~40% and ~10%, respectively, depending on the stress conditions (temperature, voltage, workload).

Designers typically add design margins to compensate for time-zero variability (due to process variation) and time-dependent (due to, e.g., bias temperature instability) variability. These variabilities become worse with scaling, which leads to larger design margin requirements. As an alternative, mitigation schemes can be applied to counteract the variability. This paper investigates the impact of aging on the offset voltage of the memory's sense amplifier (SA). For the analysis, the degradation of the SAs in the L1 data and instruction caches of an ARM processor is quantified while using realistic workloads extracted from the SPEC CPU2006 Benchmark suite. Furthermore, the effect of our mitigation scheme, i.e., an online control circuit that balances the SA workload, is analyzed. The simulation results show that the mitigation scheme reduces the offset voltage degradation due to aging with up to 40% for the benchmarks, depending on the stress conditions (temperature, voltage, and workload).

With the continuous downscaling of CMOS technologies, ICs become more vulnerable to transistor aging mainly due to Bias Temperature Instability (BTI). This paper presents a comparative study of the BTI impact while considering varying supply voltages and temperatures for three memory sense amplifier (SA) designs: low power (LP), mid power/performance (MP), and high performance (HP). As an evaluation metric, the sensing delay (SD) of the three designs is analyzed for various workloads using 45nm technology. The results show that HP SA degrades faster than MP SA and LP SA irrespective of the workload, supply voltage, and temperature. At nominal supply voltage and temperature, HP degrades up to 1.62x faster than MP, and up to 1.94x faster than LP designs for the worst case workload. In addition, the results show that an increase of 10% in power supply has a marginal impact on the relative degradation. In contrast, the results show that a temperature increment significantly worsens the BTI impact. Finally, the results show that for 16nm technology, BTI impact becomes worse and even causes read failures. This clearly indicates that designing for reliability is not only strongly application dependent, but also technology node dependent. Hence, one has to carefully consider the targeted application, design, and technology node in order to provide appropriate solutions.

This paper investigates the impact of aging in the read path of 32nm high performance SRAM; it combines the impact on the memory cell, on the sense amplifier, and on the way they interact. The analysis is done while considering different workloads and by inspecting both the bit-line swing (which reflect the degradation of the cell) and the sensing delay (which reflects the degradation of the sense-amplifier); the voltage swing on the bit lines has a direct impact on the proper functionality of the sense amplifier. The results show that in addition to the sense amplifier degradation, the cell degradation also contributes to the sensing delay increase; the share of this contribution depends on the cell design. Moreover, this sensing delay becomes worst at stressy workloads.

This paper presents litho friendly circuit techniques for variability resilient low power 8T SRAM. The new local assist circuitry achieves a state-of-the-art low energy and variability resilient WRITE operation and improves the degraded access speed of SRAM cells at low voltages. Differential VSS bias increases the variability resilience. The physical regularity in the layout of local assist circuitry enables litho optimization thereby reducing the area overhead associated with existing local assist techniques. Statistical simulations in 40nm LP CMOS technology reveals 10x reduction in WRITE energy consumption, 103x reduction in write failures, 6.5x improvement in read access time and 31% reduction in the area overhead.

This paper presents a DRAM architecture that improves the DRAM performance/power trade-off to increase their usability on low power chip design using 3D interconnect technology. The use of a finer matrix subdivision and buffering the bitline signal at the localblock level allows to reduce both the energy per access and the access time. The obtained performances match those of a typical low power SRAM, while achieving a significant area and static power reduction compared to these memories. The 128 kb memory architecture proposed here achieves an access time of 1.3 ns for a dynamic energy of less than 0.2 pJ per bit. A localized refresh mechanism allows gaining a factor of 10 in static power consumption associated with the cell, and a factor of 2 in area, when compared with an equivalent SRAM.