ArchiveMicroarchitecture design is a key stage of processor development. It is at the upper level of the entire design flow and directly affects core metrics such as performance, power consumption, and cost. Over the past few decades, new microarchitecture solutions, coupled with advances in semiconductor manufacturing, have enabled newer generations of processors to achieve higher performance, lower power consumption and cost. However, as chip design enters the post-Moore era, the dividends from the evolution of semiconductor technology are increasingly limited, and power consumption has become a major challenge for energy-efficient processor design. Meanwhile, modern processors are becoming more complex in architecture and the design space is larger, requiring designers to make accurate design metrics tradeoffs to achieve the most desirable microarchitecture design. Moreover, the existing stage-by-stage decomposition of the development and validation flow is extremely lengthy and time-consuming, and it is difficult to achieve global energy efficiency optimization. Therefore, how to perform accurate and efficient power estimation and design space exploration at the microarchitecture design stage becomes a key issue. To tackle these challenges, machine learning has been introduced into the microarchitecture design process, providing efficient and accurate solutions for microarchitecture modeling and optimization. We firstly introduce the main design flow of processors, microarchitecture design and its major challenges, then amplify machine learning-assisted integrated circuit design, which focuses on research advances in the use of machine learning techniques to assist microarchitecture power modeling and design space exploration, and finally conclude with a summary and outlook.
Fault-tolerant deep learning accelerator is the basis for highly reliable deep learning processing, and is also critical to deploy deep learning in safety-critical applications such as avionics and robotics. Since deep learning is known to be both computing-intensive and memory-intensive, traditional fault-tolerant approaches based on redundant computing will incur substantial overhead including power consumption and chip area. To this end, we propose to characterize deep learning vulnerability difference across both neurons and bits of each neuron, and leverage the vulnerability difference to enable selective protection of the deep learning processing components from the perspective of architecture layer and circuit layer respectively for the sake of lower fault-tolerant design overhead. At the same time, we observe the correlation between model quantization and bit protection overhead of the underlying processing elements of deep learning accelerators, and propose to reduce the bit protection overhead by adding additional quantization constrain without compromising the model accuracy. Finally, we employ Bayesian optimization strategy to co-optimize the correlated cross-layer design parameters at algorithm layer, architecture layer, and circuit layer to minimize the hardware resource consumption while fulfilling multiple user constraints including reliability, accuracy, and performance of the deep learning processing at the same time.
The explosive growth of the application of large-scale artificial intelligence models has made it difficult to achieve the scale deployment of applications relying on a single node or a single type of computing architecture. Distributed heterogeneous computing has become the mainstream choice, and inter-node communication has become one of the main bottlenecks in the training or inference process of large models. Currently, there are still some deficiencies in the inter-node communicating solutions dominated by leading chip manufacturers. On the one hand, some architectures choose to use a simple but less scalable point-to-point transmission scheme in order to pursue the ultimate inter-node communication performance. On the other hand, traditional heterogeneous computing engines (such as GPUs) are independent of CPUs in terms of computing resources such as memory and computing cores, but they lack dedicated communicating network devices in terms of communication resources and need to rely entirely or partially on CPUs to handle transmission between heterogeneous computing engines and the shared communicating network device through physical links such as PCIe. The proposed Direct xPU distributed heterogeneous computing architecture in this article enables heterogeneous computing engines to have independent and dedicated devices in both computing and communication resources, achieving zero-copy data and further eliminating the energy consumption and latency associated with cross-chip data transfer during inter-node communication. Evaluations show that Direct xPU achieves communication latency comparable to computing architectures pursuing ultimate inter-node communication performance, with bandwidth close to the physical limit.
As a key technique for analyzing and mining relationships, graph computing has been widely used in smart healthcare, social network analysis, financial anti-fraud, road navigation, computational sciences, and others. In recent years, with the development of parallel structures, memory access methods, interconnection structures, and synchronization mechanisms of general-purpose CPU and GPU architecture, multi-core CPU and GPU have become common platforms for accelerating graph processing. However, there are significant challenges to graph processing in terms of improving performance and achieving high scalability, such as the large scale and irregular distribution of data, complex data dependence, high communication-to-computation ratio, and dynamic changes of vertices and edges. To address the above challenges, a large number of graph processing systems based on multi-core CPU and GPU platforms are proposed and achieve good results. To provide readers with a comprehensive understanding of graph processing optimizations on multi-core CPU and GPU platforms, we first present the basic concepts, including the graph data structures, the typical graph algorithms, and the characteristics of graph application. Then the challenges of graph processing are clarified. After that, we present the existing graph processing systems based on multi-core CPU and GPU platforms. From the perspective of accelerated graph processing design, we summarize the key technology optimizations in detail and systematically, including pre-processing, memory access optimization, computing acceleration, and overhead reduction of data communication. Finally, we analyze the performance and scalability of state-of-art graph processing and conclude the future development trend of graph processing based on different perspectives, which expects to bring certain inspiration to relevant researchers to explore high-performance graph processing system.
Promoting core’s frequency is the key method for increasing performance of processor. It is hard to achieve high frequency for processor core by traditional physical design flow. Based on main place and route tools, with the same process, comparable implementation area and power consumption, our own processor core frequency can be promoted by about 30% compared with original design at signoff stage, by employing manually written block netlist, logic and physical design co-optimization, custom routing rule optimization and physical design methodology adjustment.
Processor verification occupies more than 70% of the time in the processor development flow, so it is necessary to optimize the efficiency of the processor verification process. Traditional verification methods such as software simulation provide various verification mechanisms including assertions to improve the fine-grained visibility and self-checking capability of verification, but software simulation runs slowly and lacks in efficiency. FPGA-based hardware simulation acceleration methods can greatly improve the verification performance, but debugging ability is weak, and it is difficult to locate the specific location and cause of vulnerabilities. In order to solve the above problems of verification efficiency and effectiveness, we propose a method to automatically convert non-synthesizable SystemVerilog Assertion (SVA) into logically equivalent but synthesizable RTL circuits, focusing on assertions, which is a type of non-global modeling of the design, and vertically penetrates through the various levels of abstraction, and complements the verification capability of the global ISA-based model, which can be used to verify the design. Our method complements the global ISA model-based verification capability. At the same time, combined with the advantages of FPGA fine-grained parallelization and high scalability, the verification process of the processor is hardware-accelerated, which improves the development efficiency of the processor. In this paper, we implement an end-to-end hardware assertion platform, integrate a complete toolchain for hardware-enabling SVAs, and count the triggering and coverage of hardware-enabled assertions running on FPGAs. Experiments show that the proposed method achieves more than 20000 times verification efficiency improvement compared with software simulation.
Among various cryptographic algorithms, the SM4 block cipher stands out for its simplicity and efficiency, particularly when implemented on hardware. Consequently, it has found widespread applications in encrypted transmission, encrypted storage, and beyond. As the utilization of the SM4 algorithm continues to grow, the necessity for superior hardware encryption capabilities is also increased. Recently, the implementation of the SM4 algorithm on ASIC has demonstrated high throughput in electronic code book (ECB) mode, thanks to the utilization of pipelining technology. However, in cipher block chaining (CBC) mode, achieving similar throughput improvements through pipelining is challenging due to the dependency among adjacent data blocks. To tackle this issue, we introduce two innovative simplification techniques, applied to the round function iteration process and S-box substitution process respectively. ASIC synthesis results using TSMC 40 nm technology confirm that our design achieves a throughput rate of 4.2 Gb/s in CBC mode, with a remarkable throughput of 129.4 Gb·s−1·mm−2, outperforming previously published designs in this domain.
The open-source RISC-V instruction set architecture (ISA) provides a significant opportunity for China to establish an independent and controllable IoT ecosystem. However, IoT devices often lack hardware protections and are vulnerable to physical-level firmware tampering attacks. Therefore, ensuring firmware integrity to improve device security is crucial. Previous researches have focused on secure boot techniques, but several challenges remain: 1) The traditional software-level trusted root can be physically tampered, making it difficult to ensure the reliability of the trusted root; 2) Mainstream hardware-level secure boot technology is controlled by international chip manufacturers, the specific implementation method is not publicly available, and does not support domestic cryptographic algorithms, making it impossible to ensure security independence and controllability for China; 3) Existing research on secure boot mechanisms based on RISC-V CPUs can verify the credibility of executable code but lacks a mechanism for verifying the integrity of upper-level firmware. To solve these challenges, we design and implement SeChain, a RISC-V secure boot mechanism based on the Chinese domestic cryptographic algorithm SM9. Specifically: 1) A signature calculation unit (SCU) is added inside the RISC-V SoC to generate key pairs and signatures; 2) A key verification unit (KVU) is added to perform on-chip execution of verification algorithms and firmware integrity verification; 3) A multi-level secure boot process based on verification boot is designed and implemented, starting from an immutable hardware trust root, completing the integrity verification of the next-stage boot program step by step, and achieving secure boot. Based on this design, SeChain realizes the security and trustworthiness of the trust root in the startup phase and constructs a trusted secure boot chain based on the domestic cryptographic algorithm SM9, providing reliable protection for device secure boot and trusted execution. To verify the effectiveness, efficiency, and security against trust root attacks of SeChain, we complete SeChain simulation verification experiments based on VexRiscv CPU on FPGA hardware platform. The experimental results show that SeChain can resist various firmware tampering attacks, as well as resist trust root attacks, which has an average additional time overhead of no more than 6.47 seconds. SeChain is suitable for resource constrained IoT devices, while meeting secure and trusted boot requirements, it can provide strong protection for domestic RISC-V ecosystem security independence and controllability.
With the development and successful application of deep learning, continual learning has attracted increasing attention and has been a hot topic in the field of machine learning, especially in the resource-limited and data-security scenarios with the increasing requirements of quickly learning sequential tasks and data. Unlike humans who enjoy the ability of continually learning and transferring knowledge, the existing deep learning models are prone to easily suffering from a catastrophic forgetting problem in a sequential learning process. Therefore, how to continually learn new knowledge and retain old knowledge at the same time on dynamic and non-stationary sequential task and streaming data, is the core of continual learning. Firstly, through the investigation and summary of the related work of continual learning at home and abroad in recent years, continual learning methods can be roughly divided into three categories: replay-based, constraint-based, and architecture-based. We further subdivide these three types of methods. Specifically, the replay-based methods are subdivided into three categories: sample replay, generation replay, and pseudo-sample replay, according to the sample’s sources used; the constraint-based methods are subdivided into parameter constraints, gradient constraints, and data constraints, according to the constraint’s sources; the architecture-based methods are subdivided into two categories: parameter isolation and model expansion, according to how the model structure is used. By comparing the innovation points of related work, the advantages and disadvantages of various methods are summarized. Secondly, the research progress at home and abroad is analyzed. Finally, the future development direction of continual learning combined with other fields is simply prospected.
In the era of big data, more and more data are generated in the form of data streams, which makes concept drift an important but difficult problem in streaming data mining due to its fast, infinite, unstable and dynamically changing characteristics. Most of the current concept drift processing methods have limited information extraction capability and do not fully consider the temporal features of streaming data. To address these problems, a concept drift processing method of streaming data based on mixed feature extraction (MFECD) is proposed. The method first uses convolutional kernels of different scales to model the data to construct splicing features, and uses a gating mechanism to fuse shallow inputs and splicing features for adaptive integration as different network level inputs to obtain data features that can take into account both detailed and semantic information. Based on this, attention mechanism and similarity calculation are used to evaluate the importance of stream data at different moments in order to enhance the temporal features of key site of the data stream. The experimental results show that our method can effectively extract the complex data features and temporal features contained in the streaming data, and improve the processing capability of concept drift in the data stream.
Given the excellent performance of contrastive learning on downstream tasks, contrastive clustering has received much more attention recently. However, most approaches only utilize a simple kind of data augmentation. Although augmented views keep the majority of information from original samples, they also inherit a mixture of characteristic of features, including semantic and non-semantic features, which limits model’s learning ability of semantic information under similar or identical view patterns. Even some approaches regard two different augmentation views being from the same sample and keeping similar view patterns as positive pairs, which results in sample pairs lacking of semantics. In this paper, we propose a semantic contrastive clustering method with federated data augmentation to solve these problems. Two different types of data augmentations, namely strong data augmentation and weak data augmentation, are introduced to produce two very different view patterns. These two view patterns are utilized to mitigate the disturbance of non-semantic information and improve the semantic awareness of the proposed approach. Moreover, a global k-nearest neighbor graph is used to bring global category information, which instructs the model to treat different samples from the same cluster as positive pairs. Extensive experiments on six commonly used and challenging image datasets show that the proposed method achieves the state-of-the-art performance and confirms the superiority and validity of it.
With the rapid development of Internet technology, the emergence of mobile health (mHealth) is expected to improve the quality of medical care. However, data security and user privacy issues in the mHealth field have not been fully resolved. The access control protocol based on ciphertext-policy attribute-based encryption (CP-ABE) is a promising technique for the sharing of personal health records (PHRs). However, direct adoption of the traditional CP-ABE in mHealth causes many problems. Firstly, centralized attribute authority has low ability to resist risks. Secondly, the access policies are in cleartext and leak the patient’s privacy in the encrypted PHRs. Finally, it is difficult for the traditional CP-ABE scheme to track down the user who intentionally discloses the private key. Therefore, to solve these problems, we propose a fine-grained policy-hiding and traceable decentralized access control in mHealth. This scheme implements a decentralized attribute authority mechanism. Each attribute is expressed by an attribute name and an attribute value. In the encryption phase, the attribute value is hidden in ciphertext and only generic attribute name is exposed. When the private key is maliciously leaked, the regulator can use the identity mapping table to trace the malicious user. Through experimental simulation and comparative analysis, our scheme is suitable for the actual mHealth environment in terms of security and performance.
Reversible data hiding in encrypted domain enables the secure and confidential embedding of additional information in encrypted multimedia, ensuring the privacy and integrity of both the carrier and the embedded data during transmission. The authorized recipients can extract the data without any loss and recover the media successfully. In the realm of digital media, 3D mesh models, being a relatively nascent form, possesses a distinctive file structure markedly different from that of conventional image media. Consequently, limited research has been conducted in this domain. Augmenting the embedding capacity of 3D mesh models in the encrypted domain poses an enduring challenge. The direct application of multiple most significant bit prediction algorithm from the image domain to 3D mesh models is impeded by disparities in data storage formats, thus encumbering the predictive performance of algorithms. To effectively tackle this issue, we propose the adoption of coordinate standardization to eliminate the influence of the sign bit and ameliorate the prediction algorithm’s overall performance. In order to further mitigate the inclusion of redundant auxiliary information, we introduce the integration of the selection of embedding set vertices into our experiments, which effectively generates additional payload space. The experimental results affirm that our purposed methodology attains the maximum embedding capacity while guaranteeing lossless and separable recovery of both the model and the embedded information, surpassing the capabilities of existing techniques.
Searchable symmetric encryption (SSE) has been widely concerned because of its high search efficiency. Supporting conjunctive queries can effectively improve the usability of the scheme. However, most of the existing schemes with supporting conjunctive queries just work on static databases, which do not support dynamic updates of encrypted database, such as add or delete operation. Combined with the definition of symmetric hidden vectors encryption (SHVE), a new definition, called dynamic symmetric hidden vectors encryption (DSHVE), is proposed. On the basis of this definition, a DSHVE scheme supporting update operations is constructed. At the same time, by designing oblivious dynamic cross-tags and critical data access structures based on blinded exponentiations, a dynamic searchable symmetrical encryption scheme with supporting conjunctive queries is constructed. Then details about our proposal are introduced and a formal analysis of its security is presented. Theoretic analysis and experimental results show that, without significantly increasing the storage and computing costs, only documents identifications that match the query are revealed, which avoiding revealing more result patterns. In addition, by a single round of communication for conjunctive queries, the proposed scheme has lower communication overhead and higher search efficiency.
The existing blockchain presents a multi-chain trend. Traditional consensus algorithms do not have dynamic scalability for multi-chain, making it difficult to cope with the contradiction between open use and closed maintenance of large-scale permissioned blockchain. For this problem, a novel multi-chain consensus algorithm, MC-RHotStuff, is proposed: Nodes have different roles, including alternative nodes, candidate nodes, and consensus nodes. Each working chain has consensus nodes and alternative nodes. After the admission verification, the candidate node will become an alternative node; A consensus node has a reputation value that other nodes do not have, and a consensus node that performs the correct behavior will increase the reputation value, while a consensus node that performs the wrong behavior will deduct the reputation value, then the node with abnormal reputation value will be found through the node reputation calculation and filtering algorithm MC-Scan, and a new consensus node will be selected from the alternative nodes to exchange with the abnormal node. In addition, a dynamic node adjustment algorithm, MC-Schedule, is proposed to achieve optimization by detecting the transaction volume of each blockchain and dynamically adjusting the number of consensus nodes, which not only ensures the efficient execution of the blockchain system but also improves the speed of node filtering. MC-RHotStuff proposes a node state synchronization mechanism, MC-Syn, to ensure that the consensus operates normally when the node number change or the consensus group change. Compared with existing systems, transaction throughput and latency have been comprehensively improved by about 15%.
The timely response of network services is crucial to improving user experience. Taking search engine as a typical example of network services, service providers need to ensure that the search response time is within one second. In practice, the search response time can be affected by many service attributes, such as user browsers, ISPs, and page loading methods. To optimize effectively, service providers need to identify the rules that cause high search response time, which are combinations of the above attributes. However, existing work encounters three challenges. First, the amount of search logs is large. Second, the search logs are unevenly distributed. Third, the rules with high generality are needed. Therefore, we propose a framework called Miner (multi-dimensional extraction of rules). Miner takes advantage of self-paced sampling to overcome the first and second challenges. To address the third challenge, Miner employs Corels to generate rules with high generality and recall. Our experiments use search logs provided from two top-tier search engine companies in China. The results show that Miner outperforms the state-of-the-art methods in terms of generality and recall. Operators adopt rules generated by Miner and optimize the performance of the search engine.
Remote sensing images are the data resource that can acquire information about the ocean, atmosphere, and the earth’s surface, and have been widely applied in many fields, such as agriculture, military, and urban planning. However, clouds and hazes are inevitable factors when collecting images from satellites, resulting in the loss of information and causing a huge waste of data resources in practical applications. Therefore, how to detect and remove clouds from remote sensing images is a challenging and difficult task that draws a lot of experts’ attention. We comprehensively review current research progress and summarize the challenges of cloud detection and removal in remote sensing images. Cloud detection methods are divided into two categories based on whether using deep learning technology, and cloud removal methods are divided into three categories based on whether auxiliary images are used. Then, according to the characteristics of different methods, these methods are reviewed and analyzed systematically, including their advantages and disadvantages, respectively. Afterward, four cloud detection, four thin cloud removal and four thick cloud removal methods are evaluated on two remote sensing datasets. Finally, we discuss future challenges and predict future research directions. This review paper can provide valuable advice to scientists who are involved in remote sensing image processing.