Search

Convolution features huge complexity and demands high computation capability. Among hardware platforms, Field Programmable Gate Array (FPGA) emerges as a promising solution for its substantial available parallelism and energy efficiency. Besides, convolution can be implemented with different algorithms, including conventional, GEMM, Winograd, and FFT algorithms, which are diverse in arithmetic complexity, resource requirement, etc. Different CNN models have different topologies and structures, favouring different convolution algorithms. In response, software libraries such as cuDNN provide a variety of computational primitives to support these algorithms. However, supporting such libraries on FPGAs is challenging. First, multiple algorithms can share the FPGA resources spatially as well as temporally, introducing either reconfiguration overhead or resource underutilization. Second, FPGA implementation remains a significant challenge for library developers. It typically requires significant specialized hardware knowledge. In this paper, we propose FCNNLib, an efficient and scalable convolution algorithm library on FPGAs. To coordinate multiple convolution algorithms on FPGAs, we develop three schedulings: spatial, temporal, and hybrid, which exhibit different tradeoffs in latency and throughput. We explore these schedulings by balancing the reconfiguration overhead, resource utilization and optimization objectives of the CNNs. Then, we provide efficient and tunable algorithm templates that allow performance tuning through performance and resource models. To arm the users, FCNNLib exposes a set of interfaces to support high-level application designs. We demonstrate the usability of FCNNLib with state-of-the-art CNNs. FCNNLib achieves up to 44.6X and 1.76X energy efficiency in various scenarios compared with software libraries for CPUs and GPUs, respectively.

In this research, a Graphical Processing Unit (GPU) accelerated Discrete Element Method (DEM) code was developed and coupled with the Computational Fluid Dynamic (CFD) software MFiX to simulate granular and multiphase flows with heat transfers and chemical reactions. The Fortran-based CFD solver was coupled with the CUDA/C++ based DEM solver through inter-process pipes. The speedup to the CPU version of MFiX-DEM is about 130–243 folds in the simulation of particle packings. In fluidized bed simulations, the DEM computation time is reduced from 91% to 17% with a speedup of 78 folds. The simulation of Geldart A particle fluidization revealed a similar level of importance of both fluid and particle coarse-graining. The filtered drag derived from the two-fluid model is suitable for Euler-Lagrangian simulations with both fluid and particle coarse-graining. It overcorrects the influence of sub-grid structures if used for simulations with only fluid coarse-graining.

Flexible piezocapacitive sensors utilizing nanomaterial-polymer composite based nanofibrous membranes offer an attractive alternative to more traditional piezoelectric and piezoresistive wearable sensors owing to their ultralow powered nature, fast response, low hysteresis, and insensitivity to temperature change. In this work we propose a facile method of fabricating electrospun graphene dispersed PVAc nanofibrous membrane based piezocapacitive sensors for applications in IoT enabled wearables and human physiological function monitoring. A series of electrical and material characterization experiments were conducted on both the pristine and graphene dispersed PVAc nanofibers to understand the effect of graphene addition on nanofiber morphology, dielectric response, and pressure sensing performance. Dynamic uniaxial pressure sensing performance evaluation tests were conducted on the pristine and graphene loaded PVAc nanofibrous membrane-based sensors for understanding the effect of two-dimensional (2D) nanofiller addition on pressure sensing performance. Almost twofold increase in dielectric constant and pressure sensing performance was observed for graphene loaded nanofiber sensors and subsequently, micro dipole formation model was invoked to explain the nanofiller induced dielectric constant enhancement. The robustness and reliability of the sensor has been underscored by conducting accelerated lifetime assessment experiments entailing at least 3000 cycles of periodic tactile force loading. A series of tests involving human physiological parameters monitoring were conducted to underscore the applicability of the proposed sensor for IoT enabled personalized health care, soft robotics, and next generation prosthetic devices. Finally, easy degradability of the sensing elements is demonstrated to emphasize their suitability for transient electronics applications.

The development of stimuli-responsive drug nanocarriers is an increasingly important area in nanomedicine because efficient delivery of toxic drugs to targeted tissues minimizes side effects. The specific objective of this study was to synthesize and characterize a novel pH-responsive drug carrier based on halloysite nanotubes for the controlled release of the anticancer drug doxorubicin. Poly(N,N-dimethylaminoethyl methacrylate) brushes were grafted from the surface of halloysite nanotubes using the combination of mussel-inspired polydopamine surface modification and activators regenerated by electron transfer in atom transfer radical polymerization. The chemical structure and morphology of the modified halloysite nanotubes were investigated by Fourier transform infrared spectroscopy, X-ray photoelectron spectroscopy, thermal gravimetric analysis as well as scanning and transmission electron microscopies. Dynamic light scattering and zeta potential analysis were carried out to evaluate the pH-responsivity of the functionalized halloysite nanotubes. The results of the drug loading and release study of pristine and functionalized halloysite nanotubes showed that grafting of poly(N,N-dimethylaminoethyl methacrylate) brushes on the polydopamine-modified halloysite nanotubes surface leads to a drastic increase in doxorubicin loading capacity and a highly pH-sensitive release behaviour. Less than 10 % of the loaded doxorubicin was released from poly (N,N-dimethylaminoethyl methacrylate)-grafted halloysite nanotubes at pH 7.4 after 24 h; in contrast, at pH 5.5 there was a continuous release of doxorubicin totalling 13 % in the first 30 min, i.e. lower than for the pristine halloysite nanotubes (32 %), but reaching 48 % after 24 h. Poly (N,N-dimethylaminoethyl methacrylate)-grafted halloysite nanotubes can hence be considered as a potential candidate for delivering highly toxic drug molecules to the acidic target sites.

Convolution neural networks (CNNs) as one of today’s main flavor of deep learning techniques dominate in various image recognition tasks. As the model size of modern CNNs continues to grow, neural network compression techniques have been proposed to prune the redundant neurons and synapses. However, prior techniques disconnect the software neural networks compression and hardware acceleration, which fail to balance multiple design parameters, including sparsity, performance, hardware area cost, and efficiency. More concretely, prior unstructured pruning techniques achieve high sparsity at the expense of extra performance overhead, while prior structured pruning techniques relying on strict sparse patterns lead to low sparsity and extra hardware cost. In this article, we propose OMNI, a framework for accelerating sparse CNNs on hardware accelerators. The innovation of OMNI stems from that it uses hardware amenable on-chip memory partition patterns to seamlessly engage the software CNN model compression and hardware CNN acceleration. To accelerate the compute-intensive convolution kernel, a promising hardware optimization approach is memory partition, which divides the original weight kernels into several groups so that the different hardware processing elements can simultaneously access the weight. We exploit the memory partition patterns including block, cyclic, or hybrid as a means of CNN compression patterns. Our software CNN model compression balances the sparsity across different groups and our hardware accelerator employs hardware parallelization coordinately with the sparse patterns, leading to a desirable compromise between sparsity and performance. We further develop performance models to help the designers to quickly identify the pattern factors subject to an area constraint. Last, we evaluate our design on application specific integrated circuit (ASIC) and field-programmable gate array (FPGA) platform. Experiments demonstrate that OMNI achieves $3.4\times $ – $6.2\times $ speedup for the modern CNNs, over a comparably ideal dense CNN accelerator. OMNI shows $114.7\times $ energy efficiency improvement compared with GPU platform. OMNI is also evaluated on Xilinx ZC706 and ZCU102 FPGA platforms, achieving 41.5 GOP/s and 125.3 GOP/s, respectively.

Computational fluid dynamics (CFD) is a powerful tool for prediction and analysis of complex multiphase flow hydrodynamics and residence time distribution (RTD) in chemical reactors. This study presented the validation and application of a filtered drag model for solid RTD prediction in a pilot-scale fluid catalytic cracking (FCC) circulating fluidized bed (CFB) riser with Geldart A particles. First, the filtered drag model implemented in the open-source MFiX-TFM solver was validated for flow hydrodynamics simulation in a FCC CFB riser. After that, the model was further employed to validate its ability for solid RTD prediction in the same riser by comparing the simulation results with the experimental data. The simulation with the filtered drag model well reproduced the tracer response experiment which is more accurate than that with the Gidaspow drag model. Simulations with both pulse and step tracer injection methods were compared, which reveals the limitation of solid RTD measurement using a pulse tracer injection method in the experiment.

Systolic array architecture is widely used in spatial hardware and well-suited for many tensor processing algorithms. Many systolic array architectures are implemented with high-level synthesis (HLS) design flow. However, existing HLS tools do not favor of modular and reusable design, which brings inefficiency for design iteration. In this article, we analyze the systolic array design space, and identify the common structures of different systolic dataflows. We build hardware module templates using Chisel infrastructure, which can be reused for different dataflows and computation algorithms. This remarkably improves the productivity for the development and optimization of systolic architecture. We further build a systolic array generator that transforms the tensor algorithm definition to a complete systolic hardware architecture. Experiments show that we can implement systolic array designs for different applications and dataflows with little engineering effort, and the performance throughput outperforms HLS designs.

In recent years, convolutional neural networks (CNNs) have become widely adopted for computer vision tasks. Field-programmable gate arrays (FPGAs) have been adequately explored as a promising hardware accelerator for CNNs due to its high performance, energy efficiency, and reconfigurability. However, prior FPGA solutions based on the conventional convolutional algorithm is often bounded by the computational capability of FPGAs (e.g., the number of DSPs). To address this problem, the feature maps are transformed to a special domain using fast algorithms to reduce the arithmetic complexity. Winograd and fast Fourier transformation (FFT), as fast algorithm representatives, first transform input data and filter to Winograd or frequency domain, then perform element-wise multiplication, and apply inverse transformation to get the final output. In this paper, we propose a novel architecture for implementing fast algorithms on FPGAs. Our design employs line buffer structure to effectively reuse the feature map data among different tiles. We also effectively pipeline the Winograd/FFT processing element (PE) engine and initiate multiple PEs through parallelization. Meanwhile, there exists a complex design space to explore. We propose an analytical model to predict the resource usage and the performance. Then, we use the model to guide a fast design space exploration. Experiments using the state-of-the-art CNNs demonstrate the best performance and energy efficiency on FPGAs. We achieve 854.6 and 2479.6 GOP/s for AlexNet and VGG16 on Xilinx ZCU102 platform using Winograd. We achieve 130.4 GOP/s for Resnet using Winograd and 201.1 GOP/s for YOLO using FFT on Xilinx ZC706 platform.

Knowledge on regime transition from conventional to circulating fluidization is important to scale-up and operation of liquid-solid fluidized systems in industrial applications. Previous experimental studies reported that the onset velocity measured by bed empty time test is pertinent to the physical properties of particle and liquid, and independent of solid inventory, bed geometry, and configuration. The effect of liquid properties is, however, not clear yet since tap water is usually used as the working fluid in laboratory test for convenience. To address this problem, a Computational Fluid Dynamics-Discrete Element Method (CFD-DEM) model is applied to numerically measure the onset velocity of particles in different liquid media. The model is first validated in terms of the experimental data of bed expansion in literature. Then several kinds of particles and three kinds of liquid media of different densities and viscosities are further simulated. We find that the Reynolds number based on onset velocity and Archimedes number follows a power-law relationship. With the decrease of Archimedes number, the discrepancy of Reynolds number based on onset velocity and that on particle terminal velocity becomes highly significant. The ratio of the onset velocity to particle terminal velocity for various particles in one liquid medium is not a constant, and instead dependent on both the particle and liquid properties.

Three-dimensional interconnected nanoporous graphene (NPG) microfoams and nanofoams are developed via a new approach of solid-state catalytic growth. Both NPG microfoam and nanofoam exhibit similar nanoporous structures that contain close tubular pores and open non-tubular pores but with different particle sizes. As electrochemical reactors for sulfur cathodes, sulfur is encapsulated inside the tubular pores. It is found that NPG nanoreactors can enhance the electrochemical performances in comparison with NPG microreactors, including improved reversible capacities, cyclic performances and rate performances in particular. The electrochemical impedance spectroscopy analysis reveals that NPG nanoreactors facilitate Li+ transportation and decrease the charge-transfer resistance in comparison with the microreactors, promoting the redox kinetics of multi-step conversions between sulfur and lithium sulfides. This work demonstrates a significant particle size effect of nanoporous graphene on the Li–S electrochemistry and can be useful for designing Li–S batteries as well as other electrochemical energy storage systems.

As(iii) pollution has caused increasing concern due to its significant impact in environmental safety and human health. Carbon dots (CDs)-MnO2 nanocomposites were prepared and characterized for As(iii) detection. The intense blue fluorescence of CDs can be greatly quenched by functionalization with MnO2 nanosheets due to the existence of the fluorescence resonance energy transfer (FRET) effect. CDs-MnO2 nanocomposites were then used as a fluorescence sensor for As(iii) detection with high detection sensitivity and selectivity. The redox reaction between As(iii) and MnO2 nanosheets can induce the decomposition of MnO2 and termination of the FRET process. Then the blue fluorescence originating from CDs can be recovered. The detection limit of CDs-MnO2 nanocomposites toward As(iii) was calculated to be 16.8 nM (1.40 ppb) in a linear concentration range of 0-200 nM. CDs-MnO2 nanocomposites were also found to possess highly selective ability toward As(iii) detection. In addition, the spiked and recovery test also confirmed the practicality and reliability of CDs-MnO2 nanocomposites toward As(iii) detection in real water samples, such as groundwater etc. Our research has provided a reliable tool and strategy for visual detection of As(iii) with outstanding sensing ability.

In this work, we report a class of fabric-like flexible and wearable capacitive sensors utilizing electrospun polyvinyl acetate (PVAc)-graphene nanofibers for human movement monitoring applications. The sensor reported in this work features flexible PVAc-graphene thin film sandwiched between two layers of conductive fabric and fully encapsulated in transparent adhesive films. The fabricated pressure sensor demonstrates two linear regimes with sensitivity figures of 0.0136 kPa-1 (pressure < 75 kPa) and 0.0063 kPa-1 (pressure > 75 kPa), respectively. The realized pressure sensor is flexible, apparel integrable, and the facile fabrication method will enable reliable and repeatable performance over roll-to-roll manufacturing suitable for mass production.

In recent years, attention-based models have achieved impressive performance in natural language processing and computer vision applications by effectively capturing contextual knowledge from the entire sequence. However, the attention mechanism inherently contains a large number of redundant connections, imposing a heavy computational burden on model deployment. To this end, sparse attention has emerged as an attractive approach to reduce the computation and memory footprint, which involves the sampled dense-dense matrix multiplication (SDDMM) and sparse-dense matrix multiplication (SpMM) at the same time, thus requiring the hardware to eliminate zero-valued operations effectively. Existing techniques based on irregular sparse patterns or regular but coarse-grained patterns lead to low hardware efficiency or less computation saving. This paper proposes Sanger, a framework that harvests sparsity in the attention mechanism through synergistic hardware and software co-design. The software part prunes the attention matrix into a dynamic structured pattern, and the hardware part features a reconfigurable architecture that exploits such patterns. Specifically, we dynamically sparsify vanilla attention based on a quantized prediction of the attention matrix. Then, the sparse mask is re-arranged into structured blocks that are more amenable to hardware implementation. The hardware design of Sanger features a score-stationary dataflow that keeps sparse scores stationary in the PE to avoid decoding overhead. Using this dataflow and a reconfigurable systolic array design, we can unify the computation of SDDMM and SpMM operations. Typically, the PEs can be configured during runtime to support different data access and partial sum accumulation schemes. Experiments on BERT show that Sanger can prune the model to 0.08 - 0.27 sparsity without accuracy loss, achieving 4.64X, 22.7X, 2.39X, and 1.47X speedup compared to V100 GPU, AMD Ryzen Threadripper 3970X CPU, as well as the state-of-the-art attention accelerators A3 and SpAtten.

In this work, metal foam current collectors (CCs, i.e. nickel [Ni] and copper [Cu]) were treated by thermal oxidation to create a lithiophilic oxide surface that exhibited enhanced electrochemical performances of lithium anodes such as the cyclic stability of Coulombic efficiency at different current densities for various capacities compared to pristine CCs. The oxidized CCs facilitated much increased diffusivities of ions for lithium growth than pristine CCs. It was found that an inhomogeneous solid electrolyte interphase (SEI) formed on pristine CCs while a uniform SEI formed on oxidized CCs. Uniform lithium (Li) deposition can be achieved on oxidized CCs owing to the lithiophilic oxide surface containing metal nanoparticles and the ionic compound lithium oxide (Li 2O) matrix that led to a uniform SEI film and many nucleation sites. In addition, the porous and non-porous composite anodes exhibited different electrochemical performances. The porous composite anodes showed initial lower voltage hysteresis but shorter lifetime with carbonate-based electrolyte than non-porous composite anodes. The porous composite anodes showed better rate performances in full-cell measurements while the non-porous composite anodes displayed better stability. The interfacial modification of porous hosts by lithiated oxides and the effects of porous structure on battery performances can be also useful for designing other electrodes (e.g. sodium [Na], potassium [K], zinc [Zn]).