Look-up Table Research Articles

Optimized LNS-FAID of LDPC codes: a hybrid precision decoding approach for 50G-PON.

Efficient error correction in high-speed communication networks, such as the 50G passive optical network (50G-PON), is paramount. This Letter focuses on optimizing a layered non-surjective finite alphabet iterative decoder (LNS-FAID) for 50G-PON, with an emphasis on high-throughput and low-power consumption. We propose using a distinct lookup table (LUT) for each iteration to enhance decoding performance and lower error floors. Additionally, we improve the 2-bit LNS-FAID architecture by adding operational states and a sign backtracking (SBT) strategy. This paper also introduces a hybrid precision model that merges 3-bit and 2-bit LNS-FAIDs, which balances error correction with computational efficiency. Our simulation results show that these approaches significantly improve the performance of the LDPC code in 50G-PON.

A real time hardware design for bilateral filtering

Bilateral filtering is a widely used video image filtering method, but its computational complexity is high, which is not conducive to hardware real-time applications. Therefore, this paper presents a hardware structure of bilateral filtering for real-time applications of high-definition images. Our hardware design mainly includes lookup table structure and parallel structure designs, respectively. Based on the Vivado HLS evaluation, the experimental results show that the hardware structure of this paper can meet the real-time application of bilateral filtering in high-definition video images.

Dependence of critical heat flux in vertical flow systems on dimensional and dimensionless parameters using machine learning

The critical heat flux (CHF) associated with the departure from nucleate boiling (DNB) determines the design and safety aspects of two-phase flow boiling systems. Despite the availability of several predictive tools, within the thermal engineering community, the pursuit of an accurate and robust CHF model remains a significant challenge. In this unique study, we extracted a substantial database from literature to develop machine-learning (ML) models to predict CHF for vertical flows commonly employed in the process industry. The extensive database encompasses 15,006 experimental data points gathered from diverse sources covering wide range of geometric and operational ranges of D, L, P, G, h, and x. This database is then employed to develop five ML – based models, namely Artificial Neural Networks (ANN), K-Nearest Neighbors (KNN), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Random Forests (RF). The selection of optimal input features is conducted by exploring various combinations of both dimensional and dimensionless parameters to identify the most influencing inputs for CHF prediction. The models incorporating features D, L, P, G, h, x, L/D, ρr, We, and BO achieve CHF predictions with Mean Absolute Errors (MAEs) of 8.85 % for the ANN model and 10.39 % for the XGBoost model. The optimized ML models outperformed even the highly reliable CHF correlations and lookup tables. The feature importance analysis of input parameters highlights strong dependence of CHF on dimensionless numbers such as Weber number and Boiling number.

Lookup Table-Based Design of Scalar Multiplication for Elliptic Curve Cryptography

This paper is aimed at using a lookup table method to improve the scalar multiplication performance of elliptic curve cryptography. The lookup table must be divided into two polynomials and requires two iterations of point doubling operation, for which negation operations are needed. It is well known that an inversion operation requires a lot of multiplication time. The advantage of this paper is that we are able to reduce one inverse element calculation for this problem and also improve the basic operations of finite fields through segmentation methods. If the normal basis method is used in the design of the inverse element operation, it must be converted to the normal basis through the standard basis. However, the conversion process requires a lot of matrix operations. Though the anti-element operation has good speed performance, it also increases the computational complexity. Using number theory and grouping methods will greatly improve the performance of inverse element operations. With application of the two-time point doubling operation in the hardware implementation, the developed approach reduces the computing time by 48% as compared with the conventional approach. The computational time of the scalar multiplication using the presented method is further improved by 67% over the traditional algorithm with only an area increase of 12%. Finally, the proposed lookup table-based technique can be utilized for software and hardware implementation, as the developed arithmetic operations are simple and are consistent in their execution.

Random forest evaluation using multi-key homomorphic encryption and lookup tables

In recent years, machine learning (ML) has become increasingly popular in various fields of activity. Cloud platforms have also grown in popularity, as they offer services that are more secure and accessible worldwide. In this context, cloud-based technologies emerged to support ML, giving rise to the machine learning as a service (MLaaS) concept. However, the clients accessing ML services in order to obtain classification results on private data may be reluctant to upload sensitive information to cloud. The model owners may also prefer not to outsource their models in order to prevent model inversion attacks and to protect intellectual property. The privacy-preserving evaluation of ML models is possible through multi-key homomorphic encryption (MKHE), that allows both the client data and the model to be encrypted under different keys. In this paper, we propose an MKHE evaluation method for decision trees and we extend the proposed method for random forests. Each decision tree is evaluated as a single lookup table, and voting is performed at the level of groups of decision trees in the random forest. We provide both theoretical and experimental evaluations for the proposed method. The aim is to minimize the performance degradation introduced by the encrypted model compared to a plaintext model while also obtaining practical classification times. In our experiments with the proposed MKHE random forest evaluation method, we obtained minimal (less than 0.6%) impact on the main ML performance metrics considered for each scenario, while also achieving reasonable classification times (of the order of seconds).

Five-phase space vector carrier-based PWM for third harmonic injection

Five-phase AC drives allow a third harmonic voltage to be fed into the stator winding – in addition to the fundamental voltage – to generate a third harmonic magnetic air gap field. This allows the generation of an additional small constant torque. Furthermore, the undesirable saturation harmonics that occur with saturated iron can be compensated. The desired third harmonic voltage is commonly generated via time-consuming space vector PWM algorithms, which are not well-suited for real-time application, or via look-up tables, which require an increased storage capacity. In our case, the desired third harmonic voltage is generated in the five-phase voltages by adding a non-sinusoidal zero-sequence voltage with five times the fundamental frequency to the fundamental voltage reference signals in carrier-based pulse width modulation (PWM) with a constant switching frequency. This is equivalent to a space vector PWM method (SVPWM) and avoids a significantly high calculation effort. The PWM algorithm programmed in MATLAB in real time allows the fast generation of a third harmonic voltage per phase with freely definable amplitude and phase shift without requiring a lot of storage capacity.

Impeccable Keccak

The standardization of the hash-based digital signature scheme SPHINCS+ proceeds faster than initially expected. This development seems to be welcomed by practitioners who appreciate the high confidence in SPHINCS+’s security assumptions and its reliance on well-known hash functions. However, the implementation security of SPHINCS+ leaves many questions unanswered, due to its proneness to fault injection attacks. Previous works have shown, that even imprecise fault injections on the signature generation are sufficient for universal forgery. This led the SPHINCS+ team to promote the usage of hardware countermeasures against such attacks. Since the majority of operations in SPHINCS+ is dedicated to the computation of the Keccak function, we focus on its security. At the core, hardware countermeasures against fault injection attacks are almost exclusively based on redundancy. For hash functions such as Keccak, straightforward instance- or time-redundancy is expensive in terms of chip area or latency. Further, for applications that must withstand powerful fault adversaries, these simple forms of redundancy are not sufficient. To this end, we propose our impeccable Keccak design. It is based on the methodology presented in the original Impeccable Circuits paper by Aghaie et al. from 2018. On the way, we show potential pitfalls when designing impeccable circuits and how the concept of active security can be applied to impeccable circuits. To the best of our knowledge, we are the first to provide proofs of active security for impeccable circuits. Further, we show a novel way to implement non-linear functions without look-up tables. We use our findings to design an impeccable Keccak. Assuming an adversary with the ability to flip single bits, our design detects all attacks with three and less flipped bits. Attacks from adversaries who are able to flip four or more bits are still detected with a high probability. Thus, our design is one of the most resilient designs published so far and the only Keccak design that is provably secure within a bit-flip model. At an area overhead of factor 3.2, our design is competitive with state-of-the-art designs with less resilience.

Deep-learning approach to stratified reconstructions of tissue absorption and scattering in time-domain spatial frequency domain imaging.

The conventional optical properties (OPs) reconstruction in spatial frequency domain (SFD) imaging, like the lookup table (LUT) method, causes OPs aliasing and yields only average OPs without depth resolution. Integrating SFD imaging with time-resolved (TR) measurements enhances space-TR information, enabling improved reconstruction of absorption () and reduced scattering () coefficients at various depths. To achieve the stratified reconstruction of OPs and the separation between and , using deep learning workflow based on the temporal and spatial information provided by time-domain SFD imaging technique, while enhancing the reconstruction accuracy. Two data processing methods are employed for the OPs reconstruction with TR-SFD imaging, one is full TR data, and the other is the featured data extracted from the full TR data (, continuous-wave component, , mean time of flight). We compared their performance using a series of simulation and phantom validations. Compared to the LUT approach, utilizing full TR, and datasets yield high-resolution OPs reconstruction results. Among the three datasets employed, full TR demonstrates the optimal accuracy. Utilizing the data obtained from SFD and TR measurement techniques allows for achieving high-resolution separation reconstruction of and at different depths within 5mm.

Efficient implementation of complex equations of state in a high-order framework

In the last decades, the interest in non-ideal compressible fluid dynamics (NICFD) flows and high-order accurate numerical methods, such as discontinuous Galerkin (dG), has quickly grown. In fact, advanced simulation capabilities are of paramount importance to develop new sustainable technologies with higher efficiency and low environmental impact, and to decrease the use of expensive experimental facilities. Nowadays however, the coupling of accurate Computational Fluid Dynamics (CFD) solvers with sophisticated thermodynamic models has been investigated mainly in the finite volume (FV) framework. Moreover, the solution of complex equations of state (EoS) has a too high computational cost for real-life use in industry, which is often overcome with the look-up table (LuT) interpolation approach. LuTs seem to be more suitable for FV solvers rather than high-order or finite element ones, since the interpolation error introduced may deteriorate the higher accuracy provided by these numerical methods. The novelty of this work lies in the assessment of an efficient implementation for real gas simulations in a high-order solver, by resorting to structured LuTs and automatic differentiation (AD). The proposed implementation is assessed with the computation of the inviscid and turbulent flow around a NACA0012. This approach guarantees an overhead of the computational cost around 17% with respect to an ideal gas solver with manually derived jacobian matrix.

Angiographic measurement of the superficial temporal artery for potential use in cerebral bypass surgery: a combined radiological and cadaveric study.

This study aims to investigate the microsurgical anatomy of the superficial temporal artery (STA), explore the relationship between STA length and lumen diameter, and develop a reliable radiologic method for selecting STA segments for bypass surgery. This study used 10 cadaveric dissections (20 STAs, both sides) and 20 retrospective radiological examinations (40 STAs, both sides), employing curved multiplanar reformation and flow color lookup table (CLUT) DICOM processing. Measurements included vessel lumen diameters and luminal cross-sectional thicknesses 3mm proximal to the STA bifurcation, 3mm distal to the frontal branch, 5cm distal to the frontal branch, 3mm distal to the parietal branch, and 5cm distal to the parietal branch. The distance between the STA bifurcation and the superior zygomatic border (SZB) was also measured. In our analysis, descriptive statistics encompassed mean, standard deviation (SD), standard error, minimum and maximum values, and distributions. Comparative statistics were performed using Student's t-test, with statistical significance set at p < 0.05. There were no statistically significant differences between STA measurements of bifurcation distances (p = 0.88) and lumen diameters (p = 0.46) between cadavers and radiological measures. However, lumen thicknesses were larger in frontal branches than parietal branches at the seventh and eighth centimeter (p = 0.012, p = 0.039). Branches became thinner distally from the zygoma in both cadavers and radiological image measurements. The CLUT DICOM processing radiological measures provided the high-precision required to enable pre-surgical vessel selection for extracranial-intracranial bypass. The results show that STA vessel luminal diameters are sufficient (> 1mm) for bypass surgery in the first 9cm but gradually decrease after that. Also shown is that the choice of frontal versus parietal branches depends on individual anatomical features; therefore, careful preoperative radiological examination is critical.

Target intensity correction method based on incidence angle and distance for a pulsed Lidar system.

Pulsed light detecting and ranging (Lidar) is capable of acquiring comprehensive target information within a single pulse, including distance and intensity data. Intensity data reflects the target's backscattered intensity and is commonly regarded as a crucial observational parameter associated with target reflectivity information. Multiple studies have indicated the potential of intensity data in various applications within pulsed Lidar contexts. However, the intensity data is influenced by the incident angle and distance; hence it cannot directly manifest target characteristics. Consequently, a prerequisite for its usage is the implementation of intensity calibration. This paper presents a target intensity correction method based on an improved tail model, designed for preprocessing intensity data. First, the pulse echo signal equation is derived by incorporating the improved tail model with the detected target. On this foundation, a target echo intensity correction model is established to correct the intensities at various incident angles to those at the normal direction. Lastly, the derived approach is validated through simulation analysis, and practical experiments are conducted on a constructed pulsed Lidar system. These experiments meticulously investigate the influences of incident angle and distance, two prominent factors, on echo intensity. In the context of incident angle correction experiments, the mean absolute errors (MAEs) in calibrated values for diverse targets all remain within 0.04V. Prior to correction, the maximum MAE for the cystosepiment is 0.505V; after the correction it is reduced to merely 0.02V, indicating a 96% reduction in error. Furthermore, all discrepancies exhibit an error standard deviation (ESD) of 0.03V or less, showcasing favorable stability. For distance correction, under normal incidence conditions, a diverse set of targets is measured at different distances to achieve corrected MAE and ESD within 0.05V. Consequently, the proposed method effectively achieves intensity correction concerning incident angles and distances. To achieve this, a reflectivity lookup table for the relevant targets was established. Combining this with the corrected intensity information enabled target identification in the three-dimensional imaging of pulsed Lidar.

A remote sensing algorithm for vertically resolved cloud condensation nuclei number concentrations from airborne and spaceborne lidar observations

Abstract. Cloud condensation nuclei (CCN) are mediators of aerosol–cloud interactions (ACIs), contributing to the largest uncertainties in the understandings of global climate change. We present a novel remote-sensing-based algorithm that quantifies the vertically resolved CCN number concentrations (NCCN) using aerosol optical properties measured by a multiwavelength lidar. The algorithm considers five distinct aerosol subtypes with bimodal size distributions. The inversion used the lookup tables developed in this study, based on the observations from the Aerosol Robotic Network, to efficiently retrieve optimal particle size distributions from lidar measurements. The method derives dry aerosol optical properties by implementing hygroscopic enhancement factors in lidar measurements. The retrieved optically equivalent particle size distributions and aerosol-type-dependent particle composition are utilized to calculate critical diameters using κ-Köhler theory and NCCN at six supersaturations ranging from 0.07 % to 1.0 %. Sensitivity analyses indicate that uncertainties in extinction coefficients and relative humidity greatly influence the retrieval error in NCCN. The potential of this algorithm is further evaluated by retrieving NCCN using airborne lidar from the NASA ObseRvations of Aerosols above CLouds and their intEractionS (ORACLES) campaign and is validated against simultaneous measurements from the CCN counter. The independent validation with robust correlation demonstrates promising results. Furthermore, the NCCN has been retrieved for the first time using a proposed algorithm from spaceborne lidar – Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) – measurements. The application of this new capability demonstrates the potential for constructing a 3D CCN climatology at a global scale, which helps to better quantify ACI effects and thus reduce the uncertainty in aerosol climate forcing.

Underwater image enhancement using Divide-and-Conquer network.

Underwater image enhancement has become the requirement for more people to have a better visual experience or to extract information. However, underwater images often suffer from the mixture of color distortion and blurred quality degradation due to the external environment (light attenuation, background noise and the type of water). To solve the above problem, we design a Divide-and-Conquer network (DC-net) for enhancing underwater image, which mainly consists of a texture network, a color network and a refinement network. Specifically, the multi-axis attention block is presented in the texture network, which combine different region/channel features into a single stream structure. And the color network employs an adaptive 3D look-up table method to obtain the color enhanced results. Meanwhile, the refinement network is presented to focus on image features of ground truth. Compared to state-of-the-art (SOTA) underwater image enhance methods, our proposed method can obtain the better visual quality of underwater images and better qualitative and quantitative performance. The code is publicly available at https://github.com/zhengshijian1993/DC-Net.

A multiscale computational framework for the development of spines in molluscan shells.

From mathematical models of growth to computer simulations of pigmentation, the study of shell formation has given rise to an abundant number of models, working at various scales. Yet, attempts to combine those models have remained sparse, due to the challenge of combining categorically different approaches. In this paper, we propose a framework to streamline the process of combining the molecular and tissue scales of shell formation. We choose these levels as a proxy to link the genotype level, which is better described by molecular models, and the phenotype level, which is better described by tissue-level mechanics. We also show how to connect observations on shell populations to the approach, resulting in collections of molecular parameters that may be associated with different populations of real shell specimens. The approach is as follows: we use a Quality-Diversity algorithm, a type of black-box optimization algorithm, to explore the range of concentration profiles emerging as solutions of a molecular model, and that define growth patterns for the mechanical model. At the same time, the mechanical model is simulated over a wide range of growth patterns, resulting in a variety of spine shapes. While time-consuming, these steps only need to be performed once and then function as look-up tables. Actual pictures of shell spines can then be matched against the list of existing spine shapes, yielding a potential growth pattern which, in turn, gives us matching molecular parameters. The framework is modular, such that models can be easily swapped without changing the overall working of the method. As a demonstration of the approach, we solve specific molecular and mechanical models, adapted from available theoretical studies on molluscan shells, and apply the multiscale framework to evaluate the characteristics of spines from three distinct populations of Turbo sazae.

Validating GOES Radar Estimation via Machine Learning to Inform NWP (GREMLIN) Product over CONUS

Abstract Geostationary Operational Environmental Satellites (GOES) Radar Estimation via Machine Learning to Inform NWP (GREMLIN) is a machine learning model that outputs composite reflectivity using GOES-R Series Advanced Baseline Imager (ABI) and Geostationary Lightning Mapper (GLM) input data. GREMLIN is useful for observing severe weather and initializing convection for short-term forecasts, especially over regions without ground-based radars. This study expands the evaluation of GREMLIN’s accuracy against the Multi-Radar Multi-Sensor (MRMS) System to the entire contiguous United States (CONUS) for the entire annual cycle. Regional and temporal variation of validation metrics are examined over CONUS by season, day of year, and time of day. Since GREMLIN was trained with data in spring and summer, root-mean-square difference (RMSD) and bias are lowest in the order of summer, spring, fall, and winter. In summer, diurnal patterns of RMSD follow those of precipitation occurrence. Winter has the highest RMSD because of cold surfaces mistaken as precipitating clouds, but some of these errors can be removed by applying the ABI clear-sky mask product and correcting biases using a lookup table. In GREMLIN, strong echoes are closely related to the existence of lightning and corresponding low brightness temperatures, which result in different error distributions over different regions of CONUS. This leads to negative biases in cold seasons over Washington State, lower 30-dBZ critical success index caused by high misses over the Northeast, and higher false alarms over Florida that are due to higher frequency of lightning.

Automatic smooth map generation of internal combustion engines via local-global model based calibration technique

The addition of new sensors and actuators to the engine, to reduce fuel consumption and emissions besides improving the engine operation, complicates the control commands stored in the engine control unit (ECU). Substitution of mechanical actuators with electronic ones increases the engine’s degrees of freedom and the number of control parameters, which results in the increased engine calibration time and cost. The aim of this paper is to take advantage of optimization techniques to achieve optimal values of control parameters in a fast and automated way. In this regard, it requires replacing the real engine with the virtual model and implementing the model-based calibration by coupling the virtual engine model with optimization algorithms. In this study, deep neural network (DNN) modeling and genetic algorithm (GA, NSGA-II) optimization are used for model-based calibration. The effect of all input control parameters, including ignition angle, continuously variable valve timing, etc., on all output control parameters including, brake-specific fuel consumption, emissions level, knock limit, combustion stability, etc., are investigated simultaneously by a valid global model, which is a remarkable achievement in the model-based calibration. Dynamic lag of some actuators delays the execution of control commands sent from ECU. To avoid abrupt variations in the actuators values, smoothness of the engine maps is considered in the calibration process. To reduce fuel consumption, decrease emission levels and attain smooth maps, the calibration of control parameters is performed by local-multi-objective optimization and global-single-objective optimization. Local-global model-based calibration presented in this study reduces 3.7% of the brake-specific fuel consumption and 7%–10% of emissions level at breakpoints of the engine map compared to manual calibration. In addition, the calibration time and costs while producing better engine performance can be reduced by automating the calibration process. Finally, calibrated maps are stored as a lookup table (LUT) in ECU. Generating an optimal lookup table involves the pre-calculation of several points that cover the calculation domain and allow the interpolation for other points. Selecting the optimal points for exact calculation is of great importance in the size and accuracy of LUT. In this study, an optimization tool is also presented to generate accurate and efficient LUT.

SC-IZ: A Low-Cost Biologically Plausible Izhikevich Neuron for Large-Scale Neuromorphic Systems Using Stochastic Computing

Neurons are crucial components of neural networks, but implementing biologically accurate neuron models in hardware is challenging due to their nonlinearity and time variance. This paper introduces the SC-IZ neuron model, a low-cost digital implementation of the Izhikevich neuron model designed for large-scale neuromorphic systems using stochastic computing (SC). Simulation results show that SC-IZ can reproduce the behaviors of the original Izhikevich neuron. The model is synthesized and implemented on an FPGA. Comparative analysis shows improved hardware efficiency; reduced resource utilization, which is a 56.25% reduction in slices, 57.61% reduction in Look-Up Table (LUT) usage, and a 58.80% reduction in Flip-Flop (FF) utilization; and a higher operating frequency compared to state-of-the-art Izhikevich implementation.

Enhancing FPGA Testing Efficiency: A PRBS-Based Approach for DSP Slices and Multipliers

The multiplication operations are pivotal in (Application-Specific Integrated Circuits) ASICs and Digital Signal Processors (DSPs). The integration of Field-Programmable Gate Arrays (FPGAs) into modern embedded systems, efficient Built-in Self-Tests (BISTs), particularly for complex components like DSP slices, is essential. This paper evaluates Pseudo Random Binary Sequence (PRBS) generators and checkers as BIST tools for high-speed data transfers in FPGAs. The design achieves minimal errors and remarkable efficiency with less than 4% logic utilization within available Look-Up Tables (LUTs). The testing of embedded multipliers in modern FPGAs is analyzed, shedding light on their performance. The analysis includes Built-in Self-Test (BIST), PRBS generator, PRBS checker, and Bit Error Rate (BER), providing insights into FPGA-based testing. This analysis assesses PRBS tools for high-speed FPGA data transfers. A hybrid multiplier design, featuring BIST and PRBS capabilities, notably reduces DSP slice utilization from 16% to 5%. This liberated FPGA resource enhances operational capabilities. The runtime PRBS data control at the block level design exemplifies adaptability in FPGA testing. The findings underscore PRBS-based BIST potential in FPGA testing. The hybrid multiplier not only optimizes FPGA resources but also aligns with dynamic digital system requirements. This research aids FPGA designers and engineers in advanced testing strategies for evolving embedded systems.

Improving Dust Aerosol Optical Depth (DAOD) Retrieval from the GEOKOMPSAT-2A (GK-2A) Satellite for Daytime and Nighttime Monitoring.

The Advanced Meteorological Image (AMI) onboard GEOKOMPSAT 2A (GK-2A) enables the retrieval of dust aerosol optical depth (DAOD) from geostationary satellites using infrared (IR) channels. IR observations allow the retrieval of DAOD and the dust layer altitude (24 h) over surface properties, particularly over deserts. In this study, dust events in northeast Asia from 2020 to 2021 were investigated using five GK-2A thermal IR bands (8.7, 10.5, 11.4, 12.3, and 13.3 μm). For the dust cloud, the brightness temperature differences (BTDs) of 10.5 and 12.3 μm were consistently negative, while the BTD of 8.7 and 10.5 μm varied based on the dust intensity. This study exploited these optical properties to develop a physical approach for DAOD lookup tables (LUTs) using IR channels to retrieve the DAOD. To this end, the characteristics of thermal radiation transfer were simulated using the forward model; dust aerosols were explained by BTD (10.5, 12.3 μm)-an intrinsic characteristic of dust aerosol. The DAOD and dust properties were gained from a brightness temperature (BT) of 10.5 μm and BTD of 10.5, 12.3 μm. Additionally, the cumulative distribution function (CDF) was employed to strengthen the continuity of 24-h DAOD. The CDF was applied to the algorithm by calculating the conversion value coefficient for the DAOD error correction of the IR, with daytime visible aerosol optical depth as the true value. The results show that the DAOD product can be successfully applied during the daytime and nighttime to continuously monitor the flow of yellow dust from the GK-2A satellite in northeast Asia. In particular, the validation results for IR DAOD were similar to the active satellite product (CALIPSO/CALIOP) results, which exhibited a tendency similar to that for IR DAOD at night.

Implementation of Highly Reliable Convolutional Neural Network with Low Overhead on Field-Programmable Gate Array

Due to the advantages of parallel architecture and low power consumption, a field-programmable gate array (FPGA) is typically utilized as the hardware for convolutional neural network (CNN) accelerators. However, SRAM-based FPGA devices are extremely susceptible to single-event upsets (SEUs) induced by space radiation. In this paper, a fault tolerance analysis and fault injection experiments are applied to a CNN accelerator, and the overall results show that SEUs occurring in a control unit (CTRL) lead to the highest system error rate, which is over 70%. After that, a hybrid hardening strategy consisting of a finite state machine error-correcting circuit (FSM-ECC) and a triple modular redundancy automatic hardening technique (TMR-AHT) is proposed in this paper to achieve a tradeoff between radiation reliability and design overhead. Moreover, the proposed methodology has very small workload and good migration ability. Finally, by full exploiting the fault tolerance property of CNNs, a highly reliable CNN accelerator with the proposed hybrid hardening strategy is implemented with Xilinx Zynq-7035. When BER is 2 × 10−6, the proposed hybrid hardening strategy reduces the whole system error rate by 78.95% with the overhead of an extra 20.7% of look-up tables (LUTs) and 20.9% of flip-flops (FFs).