High-performance Architecture Research Articles

Characterizing structural features of two-dimensional particle systems through Voronoi topology

Abstract This paper introduces a new approach toward characterizing local structural features of two-dimensional particle systems. The approach can accurately identify and characterize defects in high-temperature crystals, distinguish a wide range of nominally disordered systems, and robustly describe complex structures such as grain boundaries. This paper also introduces two-dimensional functionality into the open-source software program VoroTop which automates this analysis. This software package is built on a recently-introduced multithreaded version of Voro++, enabling the analysis of systems with billions of particles on high-performance computer architectures.

Towards high-performance deep learning architecture and hardware accelerator design for robust analysis in diffuse correlation spectroscopy

This study proposes a compact deep learning (DL) architecture and a highly parallelized computing hardware platform to reconstruct the blood flow index (BFi) in diffuse correlation spectroscopy (DCS). We leveraged a rigorous analytical model to generate autocorrelation functions (ACFs) to train the DL network. We assessed the accuracy of the proposed DL using simulated and milk phantom data. Compared to convolutional neural networks (CNN), our lightweight DL architecture achieves 66.7% and 18.5% improvement in MSE for BFi and the coherence factor β, using synthetic data evaluation. The accuracy of rBFi over different algorithms was also investigated. We further simplified the DL computing primitives using subtraction for feature extraction, considering further hardware implementation. We extensively explored computing parallelism and fixed-point quantization within the DL architecture. With the DL model's compact size, we employed unrolling and pipelining optimizations for computation-intensive for-loops in the DL model while storing all learned parameters in on-chip BRAMs. We also achieved pixel-wise parallelism, enabling simultaneous, real-time processing of 10 and 15 autocorrelation functions on Zynq-7000 and Zynq-UltraScale+ field programmable gate array (FPGA), respectively. Unlike existing FPGA accelerators that produce BFi and the β from autocorrelation functions on standalone hardware, our approach is an encapsulated, end-to-end on-chip conversion process from intensity photon data to the temporal intensity ACF and subsequently reconstructing BFi and β. This hardware platform achieves an on-chip solution to replace post-processing and miniaturize modern DCS systems that use single-photon cameras. We also comprehensively compared the computational efficiency of our FPGA accelerator to CPU and GPU solutions.

Elucidating the dynamics and transfer pathways of photogenerated charge carriers in V2O5/BiVO4 heterojunction photoanodes: A transient absorption spectroscopy study

Pairing of bismuth vanadate (BiVO4) with vanadium pentoxide (V2O5) forms a Type II heterojunction photoanode, which has been proven to be a high-performance photoanode architecture for efficient photoelectrochemical (PEC) water oxidation. To further support for advanced rational design and improvement of the heterojunction photoanode, it is quintessential to understand the mechanics and properties of photogenerated charge carriers formed. This study aims to probe the dynamics of photogenerated electron-hole pairs formed in pristine photoanode as well as the heterojunction photoanodes using transient absorption spectroscopy (TAS). Relative to the BiVO4/V2O5 structure, modelling and quantification of the decay constant helps in explaining why the V2O5/BiVO4 heterojunction photoanode exhibited a longer lifetime of the photogenerated charge carriers with a lower decay rate constant that leads to a much-improved overall generation of photocurrent density. An oxygen evolving catalyst of nickel oxyhydroxide (NiOOH) was then anchored on the exterior surfaces of the heterojunction photoanode system to further investigate and modify the transfer pathways of the photogenerated charge carriers. Furthermore, the hole- and scavenger-assisted TAS measurements on the BiVO4/V2O5 heterojunction photoanode revealed that the majority of trapped holes species are accumulated within the BiVO4 layer. The findings suggested that the Type II heterojunction formation in V2O5/BiVO4 could effectively reduce its charge recombination process.

OStr-DARTS: Differentiable Neural Architecture Search Based on Operation Strength.

Differentiable architecture search (DARTS) has emerged as a promising technique for effective neural architecture search, and it mainly contains two steps to find the high-performance architecture. First, the DARTS supernet that consists of mixed operations will be optimized via gradient descent. Second, the final architecture will be built by the selected operations that contribute the most to the supernet. Although DARTS improves the efficiency of neural architecture search (NAS), it suffers from the well-known degeneration issue which can lead to deteriorating architectures. Existing works mainly attribute the degeneration issue to the failure of its supernet optimization, while little attention has been paid to the selection method. In this article, we cease to apply the widely-used magnitude-based selection method and propose a novel criterion based on operation strength that estimates the importance of an operation by its effect on the final loss. We show that the degeneration issue can be effectively addressed by using the proposed criterion without any modification of supernet optimization, indicating that the magnitude-based selection method can be a critical reason for the instability of DARTS. The experiments on NAS-Bench-201 and DARTS search spaces show the effectiveness of our method.

High-Performance Method and Architecture for Attention Computation in DNN Inference.

In recent years, The combination of Attention mechanism and deep learning has a wide range of applications in the field of medical imaging. However, due to its complex computational processes, existing hardware architectures have high resource consumption or low accuracy, and deploying them efficiently to DNN accelerators is a challenge. This paper proposes an online-programmable Attention hardware architecture based on compute-in-memory (CIM) marco, which reduces the complexity of Attention in hardware and improves integration density, energy efficiency, and calculation accuracy. First, the Attention computation process is decomposed into multiple cascaded combinatorial matrix operations to reduce the complexity of its implementation on the hardware side; second, in order to reduce the influence of the non-ideal characteristics of the hardware, an online-programmable CIM architecture is designed to improve calculation accuracy by dynamically adjusting the weights; and lastly, it is verified that the proposed Attention hardware architecture can be applied for the inference of deep neural networks through Spice simulation. Based on the 100nm CMOS process, compared with the traditional Attention hardware architectures, the integrated density and energy efficiency are increased by at least 91.38 times, and latency and computing efficiency are improved by about 12.5 times.

Correction: High-performance visual geometric group deep learning architectures for MRI brain tumor classification

Correction: High-performance visual geometric group deep learning architectures for MRI brain tumor classification

AutoAMS: Automated attention-based multi-modal graph learning architecture search

Multi-modal attention mechanisms have been successfully used in multi-modal graph learning for various tasks. However, existing attention-based multi-modal graph learning (AMGL) architectures heavily rely on manual design, requiring huge effort and expert experience. Meanwhile, graph neural architecture search (GNAS) has made great progress toward automatically designing graph-based learning architectures. However, it is challenging to directly adopt existing GNAS methods to search for better AMGL architectures because of the search spaces that only focus on designing graph neural network architectures and the search objective that ignores multi-modal interactive information between modalities and long-term content dependencies within different modalities. To address these issues, we propose an automated attention-based multi-modal graph learning architecture search (AutoAMS) framework, which can automatically design the optimal AMGL architectures for different multi-modal tasks. Specifically, we design an effective attention-based multi-modal (AM) search space consisting of four sub-spaces, which can jointly support the automatic search of multi-modal attention representation and other components of multi-modal graph learning architecture. In addition, a novel search objective based on an unsupervised multi-modal reconstruction loss and task-specific loss is introduced to search and train AMGL architectures. The search objective can extract the global features and capture multi-modal interactions from multiple modalities. The experimental results on multi-modal tasks show strong evidence that AutoAMS is capable of designing high-performance AMGL architectures.

A novel high-performance TG-based SRAM cell with 5 nm FinFET technology

In this study, we investigate the performance and reliability of a novel static random-access memory (SRAM) cell utilizing advanced 5 nm FinFET technology. Our research aims to address critical challenges in SRAM design by integrating transmission gates and power gated transistors. Through extensive simulations using the Cadence Virtuoso tool, we optimize the SRAM cell’s read and write paths, resulting in substantial improvements in both functionalities. Additionally, our study unveils temperature-dependent variations in the read current and write margin, emphasizing the influence of temperature on SRAM performance. Compared to conventional FinFET SRAM circuits of equivalent bit-cell area and read latency, our innovative design showcases remarkable improvements across various parameters. Specifically, we achieve a commendable increase of 6.16% in the write static noise margin (WSNM) and 5.86% in the hold static noise margin (HSNM). Moreover, our findings reveal a substantial boost in read stability, increasing from 14.75% to 18.35%. These advancements underscore the promising potential of our approach in paving the way for future innovations in high-performance memory architectures. By leveraging state-of-the-art technology and meticulous optimization techniques, our research sets a new standard for SRAM design, offering enhanced performance, reliability, and efficiency in memory systems.

Hydrothermal synthesis of high-performance cobalt phosphate nanorod architecture as bifunctional electrocatalysts for rechargeable Zn-air batteries and supercapatteries

Transition metal phosphate bifunctional electrocatalysts demonstrate superior oxygen evolution reaction (OER), oxygen reduction reaction (ORR), and capacitive properties owing to their high electrocatalytic and conductive activities. In this report, cobalt phosphate (Co3P2O8) nanorod (NR) electrocatalytic materials were developed via a facile one-step hydrothermal synthesis, followed by an annealing process. The unique NR structures offer superior electrocatalytic activities, continuous and quick electron transport pathways, short ion diffusion lengths, and rich active sites toward rechargeable zinc-air batteries (ZABs) and supercapatteries (SCs). The self-assembled Co3P2O8 NR electrocatalyst revealed higher current density values during the OER and ORR measurements than the commercially available Pt/C and IrO2 materials. The Co3P2O8 NR electrocatalyst exhibited an excellent OER overpotential of 180 mV. The electocatalyst also showed superior onset and half-wave potentials of 0.92 V and 0.80 V, respectively. Moreover, the fabricated Co3P2O8 NRs-based ZAB exhibited better durability than the Pt/C-based ZAB. Significantly, the Co3P2O8 NR electrode demonstrated an excellent specific capacity of 373 mAh g−1 with outstanding cycling stability of 99% after 5000 cycles. Additionally, the fabricated Co3P2O8 NRs//activated carbon SC demonstrated high specific energy and power densities of 25.62 Wh kg−1 and 2953.12 W kg−1, respectively with remarkable cycling stability of 89% after 7000 cycles. Specifically, the practical applicability of the fabricated Co3P2O8 NRs-based ZAB and SC was tested using a blue light-emitting diode nameplate energized by two devices in series. Overall, the catalytic and energy-storage properties of the Co3P2O8 NR material can provide a new pathway for high-energy electrochemical applications.

A Generic High-Performance Architecture for VPN Gateways

Virtual private network (VPN) gateways are widely applied to provide secure end-to-end remote access and to relay reliable interconnected communication in cloud computing. As network convergence nodes, the performance of VPN gateways is limited by traditional methods of packet receiving and sending, the kernel protocol stack and the virtual network interface card. This paper proposes a generic high-performance architecture (GHPA) for VPN gateways in consideration of its generality and performance. In terms of generality, we redesign a generic VPN core framework by modeling a generic VPN communication model, formulating generic VPN core technologies and presenting corresponding core algorithms. In terms of performance, we propose a three-layer GHPA for VPN gateways by designing a VPN packet processing layer based on a data plane development kit (DPDK), implementing a user space basic protocol stack and applying our proposed generic VPN core framework. On the basis of the research work above, we implement a high-performance VPN (HP-VPN) and a traditional VPN (T-VPN) that complies with GHPA and traditional methods, respectively. Experimental results prove that the performance of HP-VPN based on GHPA is superior to T-VPN and other common VPNs in RTT, system throughput, packet forwarding rate and jitter. In addition, GHPA is extensible and applicable for other VPN gateways to improve their performance.

3D Architecting triple gradient graphene-based fiber electrode for high-performance asymmetric supercapacitors

High-performance fibrous architectures and controllable fabrication methods for active nanomaterials are fundamental in the development of fiber-based supercapacitors (FSCs). Herein, we rationally designed a triple gradient structured fibrous electrode based on three-dimensional (3D) architectural approach with Ni3S2 nanoflakes (Ni3S2 NFs) in situ growth on Ni-coated graphene fiber (Ni3S2-NiGFs) for highly energy-dense asymmetric supercapacitors. Remarkably, the Ni-metal coating imparted the Ni3S2-NiGFs exceptional electrical conductivity (619.8 S cm−1), mechanical strength (335.0 MPa), and flexibility without significantly compromising its density. Moreover, the unique 3D interconnected ultrathin Ni3S2 NFs significantly increase the contact area between the electrode and the electrolyte solution, providing abundant redox activity to facilitate ion transport and accumulation. Therefore, the fibrous electrode showcased a remarkable areal specific capacitance of 1387.6 mF cm−2 at 3 mA cm−2 in 3 M KOHaq electrolyte, while its asymmetric FSCs delivered an impressive areal energy density of 15.6 μW h cm−2 at power density of 1.0 mW cm−2 along with durable cycle life (86.7 % of the initial specific capacitance retention over 10,000 cycles). In summary, the featured work provides a facile but robust electrode design to realize highly energy-dense flexible energy storage devices for next-generation wearable industries.

Thermal comfort, daylight, and energy performance of envelope-integrated algae-based bioshading and static shading systems through multi-objective optimization

Energy-efficient envelopes integrated with different shading devices, as noticeable passive design strategies, have been of great interest for research studies. However, there is a lack of attention on optimizing envelope-integrated novel algae-based bioshading systems (ABBS) in comparison with other shading devices to enhance buildings’ sustainability-oriented performances. Integrating microalgae culture system with building façade as a state-of-the-art bioshading system is among recent developments in high-performance architecture. Although, challenges/limitations concerning different performances of this technology have not been extensively addressed. The present study conducts a multi-objective optimization framework to investigate the Thermal Comfort Percentage (TCP), Useful Daylight Illuminance (UDI), and Energy Usage Intensity (EUI) in a school building to evaluate the relationship between the performance objectives and design variables (shading characteristics, window-to-wall ratio (WWR)) through comparing different static shading systems and ABBS in the hot (BSh) and cold semi-arid (BSk) climates. The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) method was also applied to determine the best design options within Pareto frontiers. Results demonstrated that horizontal louver by 128.52 % TCP, 15.10 % UDI improvement in south-oriented façade in BSk, and overhang with fin (OF) by 18.99 % EUI reduction in southern facade in BSh climates contribute the most to enhance the objective metrics. Also, ABBS was not a stand-first system to enhance all the objective metrics compared to the other shading systems in none of the examined climates. Sensitivity analysis indicated that WWR has the most significant positive impact on EUI compared to the other studied metrics in almost all design options.

The optimization of roll-bond cold plate based on Taguchi-grey theory for server chips thermal management

The roll-bond cold plate, an economical and straightforward electronics liquid cooling approach, is highly promising to address the increasing electronics chip thermal management challenges. However, the roll-bond cold plate with staggered cellular flow channels, while having a sizable solid-liquid contact area and high heat dissipation potential, suffers from inadequate temperature uniformity and fluidity. This study addresses flow and temperature uniformity issues in current roll-bond cold plates by proposing a design—Direct Liquid Cooling Roll-Bond Plates (DLCRBP) tailored for high-performance server architectures. To investigate how regional combination, cellular diameter, and longitudinal spacing affect both thermal and flow performance, orthogonal experiments with diverse flow channel structures were conducted using the Taguchi-grey theory. COMSOL-based numerical simulation were used to analyze these experiment conditions. Multi-objective optimization, considering pressure drop, heat source temperature uniformity, and coolant temperature difference, was then performed using the grey correlation method. The optimized DLCRBP (A3B5C1) was fabricated, and its performance was compared with the staggered cellular DLCRBP (SCS), which is the control group without optimization, under various inlet flow velocities, heat source powers, and coolant inlet temperatures. Experimental comparisons illustrate that the optimized structure ensures outstanding temperature uniformity even at low inlet flow velocities. Specifically, at an inlet flow velocity of 0.2 m/s, the average and maximum temperature differences on the heat source's upper surface are 36.5 % and 75.3 % lower, respectively, compared to pre-optimization levels. Furthermore, A3B5C1's temperature uniformity improves with higher coolant inlet flow velocity and lower coolant inlet temperatures. At a coolant inlet flow velocity of 0.5 m/s or higher, A3B5C1 sustains the surface temperature of a 600 W chip below the thermal design's safe temperature of 77.8 °C, highlighting its enhanced performance.

Damascene versus subtractive line CMP process for resistive memory crossbars BEOL integration

In recent years, resistive memories have emerged as a pivotal advancement in the realm of electronics, offering numerous advantages in terms of energy efficiency, scalability, and non-volatility [1]. Characterized by their unique resistive switching behavior, these memories are well-suited for a variety of applications, ranging from high-density data storage to neuromorphic computing [2]. Their potential is further enhanced by their compatibility with advanced semiconductor processes, enabling seamless integration into modern electronic circuits [3]. A particularly promising avenue for resistive memory lies in its integration at the Back-End-of-Line (BEOL) stage of semiconductor manufacturing [4]. BEOL integration involves processes that occur after the fabrication of the transistors, primarily focusing on creating interconnections that electrically link these transistors. Integrating resistive memories at this stage can lead to compact, efficient, and high-performance architectures, pivotal for in-memory computing applications where data storage and processing are co-located [5]. This paper studies three ways to integrate TiOx-based resistive memory into passive crossbar array structures, using chemical mechanical polishing (CMP) processes, focusing on identifying the optimal integration techniques.

OnceNAS: Discovering efficient on-device inference neural networks for edge devices

Edge Intelligence (EI) offers an attractive approach for local AI processing at the network edge for privacy protection and reduced transmission, but deploying resource-intensive neural networks on edge devices remains a challenge. The neural architecture search (NAS) technique, known for its automation and minimal manual intervention, serves as a pivotal tool for EI. However, existing methods typically concentrate on optimizing resource consumption for specific hardware, leading to hardware-specific neural architectures with limited generalizability. In response, we propose OnceNAS, a novel method that designs and optimizes on-device inference neural networks for resource-constrained edge devices. OnceNAS simultaneously optimizes for parameter count and inference latency in addition to inference accuracy, producing lightweight neural networks while maintaining their inference performance. Meanwhile, we introduce an efficient evaluation strategy that can simultaneously assess multiple metrics. Experimental results demonstrate the effectiveness of OnceNAS, achieving high-performing architectures with substantial size reduction (10.49x) and speedup (5.45x). As a result, OnceNAS offers practical value by generating efficient on-device inference neural architectures for resource-constrained edge devices, facilitating real-world applications like autonomous driving and smart healthcare. Furthermore, we contribute DARTS-Bench, an open-source dataset providing candidate architectures with hardware-related information and a user-friendly API, facilitating future research in lightweight NAS.

Swift : a modern highly parallel gravity and smoothed particle hydrodynamics solver for astrophysical and cosmological applications

ABSTRACT Numerical simulations have become one of the key tools used by theorists in all the fields of astrophysics and cosmology. The development of modern tools that target the largest existing computing systems and exploit state-of-the-art numerical methods and algorithms is thus crucial. In this paper, we introduce the fully open-source highly-parallel, versatile, and modular coupled hydrodynamics, gravity, cosmology, and galaxy-formation code Swift. The software package exploits hybrid shared- and distributed-memory task-based parallelism, asynchronous communications, and domain-decomposition algorithms based on balancing the workload, rather than the data, to efficiently exploit modern high-performance computing cluster architectures. Gravity is solved for using a fast-multipole-method, optionally coupled to a particle mesh solver in Fourier space to handle periodic volumes. For gas evolution, multiple modern flavours of Smoothed Particle Hydrodynamics are implemented. Swift also evolves neutrinos using a state-of-the-art particle-based method. Two complementary networks of sub-grid models for galaxy formation as well as extensions to simulate planetary physics are also released as part of the code. An extensive set of output options, including snapshots, light-cones, power spectra, and a coupling to structure finders are also included. We describe the overall code architecture, summarize the consistency and accuracy tests that were performed, and demonstrate the excellent weak-scaling performance of the code using a representative cosmological hydrodynamical problem with ≈300 billion particles. The code is released to the community alongside extensive documentation for both users and developers, a large selection of example test problems, and a suite of tools to aid in the analysis of large simulations run with Swift.

Swin transformer network leveraging multi-dimensional features for defect depth prediction

This study introduces a novel method for accurately predicting defect depth in pulsed infrared thermography. The core innovation of this study lies in the utilization of multi-dimensional features to enhance the accuracy of depth prediction. By integrating temperature, temperature change rate, and time-frequency spectrum into a comprehensive feature set, we aim to capture a more detailed understanding of defect characteristics, thereby facilitating more precise predictions. Additionally, we employ the Hilbert encoding method to obtain two-dimensional matrices and utilize mask-based augmentation to generate synthetic defect data matrices. Subsequently, the generated matrices are fed input into a Swin transformer network configured with shifted windows and multi-head self-attention. In comparison to existing high-performing architectures, our method leveraging multi-dimensional features, demonstrates superior performance, especially in defect depth prediction for non-planar samples. This work paves the way for more accurate and efficient defect depth prediction in various infrared thermography applications.

A universal parallel simulation framework for energy pipeline networks on high-performance computers

Energy distribution networks represent crucial infrastructures for modern society, and various simulation tools have been widely used by energy suppliers to manage these intricate networks. However, simulation calculations include a large number of fluid control equations, and computational overhead limits the performance of simulation software. This paper proposes a universal parallel simulation framework for energy pipeline networks that takes advantages of data parallelism and computational independence between network elements. A non-pipe model of an energy supply network is optimized, and the input and output of the network model in the proposed framework are modified, which can reduce the development burden during the numerical computations of the pipeline network and weaken the computational correlation between different simulated components. In addition, independent computations can be performed concurrently through periodic data exchange procedures between component instances, improving the parallelism and efficiency of simulation computations. Further, a parallel water pipelines network simulation computing paradigm based on a heterogeneous computer hardware architecture is used to evaluate the proposed framework’s performance. A series of tests are conducted to verify the accuracy of the proposed framework, and simulation errors of less than 5% are achieved. The results of multi-threaded simulation experiments have demonstrated the feasibility of the proposed framework in a parallel computing approach. Moreover, an Advanced Micro Devices (AMD) Deep Computing Unit (DCU)-parallel program is implemented into a water supply network simulation system; the computational efficiency of this system is compared with that of its serial counterpart. The experimental results show that the proposed framework is appropriate for high-performance computer architectures, and the 18x speed-up ratio demonstrates that the parallel program based on the proposed universal framework outperforms the serial program. That provides the basis for the application of pipe network simulation on high-performance computers.

Enhancing Interface Connectivity for Multifunctional Magnetic Carbon Aerogels: An In Situ Growth Strategy of Metal-Organic Frameworks on Cellulose Nanofibrils.

Improving interface connectivity of magnetic nanoparticles in carbon aerogels is crucial, yet challenging for assembling lightweight, elastic, high-performance, and multifunctional carbon architectures. Here, an in situ growth strategy to achieve high dispersion of metal-organic frameworks (MOFs)-anchored cellulose nanofibrils to enhance the interface connection quality is proposed. Followed by a facile freeze-casting and carbonization treatment, sustainable biomimetic porous carbon aerogels with highly dispersed and closely connected MOF-derived magnetic nano-capsules are fabricated. Thanks to the tight interface bonding of nano-capsule microstructure, these aerogels showcase remarkable mechanical robustness and flexibility, tunable electrical conductivity and magnetization intensity, and excellent electromagnetic wave absorption performance. Achieving a reflection loss of -70.8dB and a broadened effective absorption bandwidth of 6.0GHz at a filling fraction of merely 2.2wt.%, leading to a specific reflection loss of -1450dBmm-1, surpassing all carbon-based aerogel absorbers so far reported. Meanwhile, the aerogel manifests high magnetic sensing sensibility and excellent thermal insulation. This work provides an extendable in situ growth strategy for synthesizing MOF-modified cellulose nanofibril structures, thereby promoting the development of high-value-added multifunctional magnetic carbon aerogels for applications in electromagnetic compatibility and protection, thermal management, diversified sensing, Internet of Things devices, and aerospace.

Optimizing vision transformers for CPU platforms via human-machine collaborative design

Over the past three years, various vision transformers (ViTs) have been proposed and applied to many vision tasks. However, for x86 CPU platforms, current ViTs fail to effectively balance inference efficiency and accuracy because they are specifically designed for high-end GPU platforms or smart mobile platforms. In this paper, we aim to explore high-performance transformer architectures for CPU platforms via human-machine collaborative design. First, we attempt to discover CPU-unfriendly design choices in existing ViTs through several controlled experiments. According to our observations, we then design two CPU-efficient attention units for building transformers by human manual exploration. Finally, based on our basic units, we present a CPU-Aware network search method to automatically search for global architectures of transformers. Through the process of human-machine collaborative design, we can obtain ViT instances. Benefiting from our human-designed basic units and automatically-searched global structures, our ViT instances achieve a more favorable tradeoff between CPU-latency and accuracy. For CPUs based on x86 architecture, they are more competitive than their CNN-based and transformer-based counterparts. This suggests that our proposed human-machine collaborative solution is feasible for designing vision transformers tailored for specific hardware.