Transformer-based large language models (LLMs) such as Generative Pre-trained Transformer (GPT) have become popular due to their remarkable performance across diverse applications, including text generation and translation. For LLM training and inference, the GPU has been the predominant accelerator with its pervasive software development ecosystem and powerful computing capability. However, as the size of LLMs keeps increasing for higher performance and/or more complex applications, a single GPU cannot efficiently accelerate LLM training and inference due to its limited memory capacity, which demands frequent transfers of the model parameters needed by the GPU to compute the current layer(s) from the host CPU memory/storage. A GPU appliance may provide enough aggregated memory capacity with multiple GPUs, but it suffers from frequent transfers of intermediate values among GPU devices, each accelerating specific layers of a given LLM. As the frequent transfers of these model parameters and intermediate values are performed over relatively slow device-to-device interconnects such as PCIe or NVLink, they become the key bottleneck for efficient acceleration of LLMs. Focusing on accelerating LLM inference, which is essential for many commercial services, we develop CXL-PNM, a processing near memory (PNM) platform based on the emerging interconnect technology, Compute eXpress Link (CXL). Specifically, we first devise an LPDDR5X-based CXL memory architecture with 512GB of capacity and 1.1TB/s of bandwidth, which boasts 16× larger capacity and 10× higher bandwidth than GDDR6and DDR5-based CXL memory architectures, respectively, under a module form-factor constraint. Second, we design a CXLPNM controller architecture integrated with an LLM inference accelerator, exploiting the unique capabilities of such CXL memory to overcome the disadvantages of competing technologies such as HBM-PIM and AxDIMM. Lastly, we implement a CXLPNM software stack that supports seamless and transparent use of CXL-PNM for Python-based LLM programs. Our evaluation shows that a CXL-PNM appliance with 8 CXL-PNM devices offers 23% lower latency, 31% higher throughput, and 2.8× higher energy efficiency at 30% lower hardware cost than a GPU appliance with 8 GPU devices for an LLM inference service.

Emerging applications such as deep neural network demand high off-chip memory bandwidth. However, under stringent physical constraints of chip packages and system boards, it becomes very expensive to further increase the bandwidth of off-chip memory. Besides, transferring data across the memory hierarchy constitutes a large fraction of total energy consumption of systems, and the fraction has steadily increased with the stagnant technology scaling and poor data reuse characteristics of such emerging applications. To cost-effectively increase the bandwidth and energy efficiency, researchers began to reconsider the past processing-in-memory (PIM) architectures and advance them further, especially exploiting recent integration technologies such as 2.5D/3D stacking. Albeit the recent advances, no major memory manufacturer has developed even a proof-of-concept silicon yet, not to mention a product. This is because the past PIM architectures often require changes in host processors and/or application code which memory manufacturers cannot easily govern. In this paper, elegantly tackling the aforementioned challenges, we propose an innovative yet practical PIM architecture. To demonstrate its practicality and effectiveness at the system level, we implement it with a 20nm DRAM technology, integrate it with an unmodified commercial processor, develop the necessary software stack, and run existing applications without changing their source code. Our evaluation at the system level shows that our PIM improves the performance of memory-bound neural network kernels and applications by 11.2× and 3.5×, respectively. Atop the performance improvement, PIM also reduces the energy per bit transfer by 3.5×, and the overall energy efficiency of the system running the applications by 3.2×.

Sensory information from an Earth-fixed reference is necessary to guarantee a high localization accuracy of an unmanned underwater vehicle (UUV). However, the implementation of these sensors in an underwater environment is challenging because of signal uncertainties and strong signal attenuation. In this paper, we propose an underwater localization scheme with a sensor fusion of inertial navigation system (INS) and received signal strength of electromagnetic (EM) waves sensors. In the proposed sensor-fusion-based localization scheme, the UUV predicts its location by using INS based on dead-reckoning and corrects the predicted position by Kalman filter using EM waves sensor information when the UUV receives the signals of EM waves sensors in underwater wireless sensor networks. The proposed scheme enables localization with high accuracy and high sampling rate during a long-term task. The results of an experiment performed in a basin environment shows the feasibility of the proposed scheme. The scheme achieved reliable localization accuracy by comparing the pre-measured ground-truth position and long-term navigation. These results show the feasibility of exploiting EM waves attenuation as Earth-fixed reference sensors.

Several wavelet synopsis construction algorithms were previously proposed based on dynamic programming for unrestricted Haar wavelet synopses as well as Haar + synopses. However, they find an optimal synopsis for every incoming value in each node of a coefficient tree, even if different incoming values share an identical optimal synopsis. To alleviate the limitation, we present novel algorithms, which keep only a minimal set of the distinct optimal synopses in each node of the tree, for the error-bounded synopsis problem. Furthermore, we propose the methods to restrict coefficient values to be considered to compute the optimal synopses in each node. In addition, by partitioning all optimal synopses in each node into a set of groups, such that every group can be represented by a compact representation, we significantly improve the performance of the proposed algorithms.

Measuring similarity between source codes has lots of applications, such as code plagiarism detection, code clone detection, and malware detection. A variety of methods for the measurement have been developed and program-dependence-graph based methods are known to be well working against disguise techniques. But these methods usually rely on solving NP-hard problems which cause a scalability issue. In this paper, we propose a genetic algorithm to measure the similarity between two codes by solving an error correcting subgraph isomorphism problem on dependence graphs. We propose a new cost function for this problem, which reflects the characteristic of the source codes. An incremental genetic algorithm is used to solve the problem. The size of the graph to be searched is gradually increasing during the evolutionary process. We developed new operators for the algorithm, and the overall system is tested on some real world data. Experimental results showed that the system successfully works on code plagiarism detection and malware detection. The similarity computed by the system turned out to reflect the similarity between the codes properly.

Previously, we proposed a scheme that determines the position of a remotely operated underwater vehicle (ROV) from the signal strengths of commercial radio-frequency sensors and antennas. This scheme provides accurate position information in a structured environment but is limited to two-dimensional (2D) environments because the radiation power of the antenna depends on the elevation angle between the sending and receiving antennas. To overcome this problem, we propose a 3D localization scheme that considers the electromagnetic (EM) wave attenuation over the range of reliable elevation angles. In order to determine the reliable elevation scope, we analyzed the radiation patterns of dipole antennas. The feasibility of our approach is demonstrated in distance estimation and 3D localization experiments by varying the distance and elevation angle. Encouraged by these results, we constructed an underwater wireless sensor network in the experimental basin, and performed ROV position tracking with the depth sensor. The scheme achieved reliable localization accuracy at a fast sampling rate, demonstrating the feasibility of exploiting EM wave attenuation in localization.

In this paper, we propose a genetic algorithm for a LEGO(R) brick layout problem. The task is to build a given 3D object with LEGO(R) bricks. A brick layout is modeled as a solution to a combinatorial optimization problem, through intermediate voxelization, which tries to maximize the connectivity and then minimize the number of used bricks. We attack the problem in the context of genetic search. The proposed randomized greedy algorithm produces initial solutions, and the solutions are effectively improved by an evolutionary process. New domain-specific methods are proposed as well, which include a random boundary mutation and a thickening approach. We tested our algorithm on various objects collected from the web. Experimental results showed that the algorithm produces efficient, and mostly optimal solutions for benchmark models. Unlike some previous works, our algorithm is not limited to assemble few specific objects, but it can deal with diverse kind of objects. To the best of our knowledge, this is the most extensive empirical study on the problem.

We propose a scheme that uses the signal strength received from commercial RF sensors to determine the position of a sensor underwater. The scheme can give accurate position information and has fast update rate, but is limited to short sensor range with the signal identification problem. The infrastructure-based localization and a method that assigns frequency bands to RF sensors can be a solution to overcome this range limitation without loss of update rate or position accuracy. The RF sensor modules are used as ANs, and are easily identified by their channel assignments. To verify the proposed localization scheme, we organized an experimental environment and performed repetitive experiments to track the positions of mobile nodes and stationary nodes in 2D environments. The experimental results showed reliable localization accuracy with fast sampling rate and demonstrates the feasibility of localization that exploits EM wave attenuation.

Finding an isomorphic subgraph is a key problem in many real world applications modeled on graph. In this paper, we propose a new hybrid genetic algorithm(GA) for subgraph isomorphism problem which uses an incremental approach. We solve the problem with increasing the size of the subproblem step by step. The graph for which we search is gradually expanded from the empty structure to the entire one. We apply a hybrid GA to each subproblem, initialized with the evolved population of previous step. We present design issues for the incremental approach, and the effects of each design decision are analyzed by experiment. The proposed algorithm is tested on widely used dataset. With apposite vertex reordering along with moderate population diversity, incremental approach brought a significant performance improvement. Experimental results showed that our algorithm outperformed representative previous works.

In this paper, we discuss a novel underwater localization system using EM(Electro-Magnetic) wave. Maxwell equation and Friis transmission formula were reconstructed in terms of propagation constant and antenna property, then by combining these two formulas, we have defined the EM wave attenuation function over underwater distance. Using this function, we have developed an underwater sensor model for EM waves. Finally we organized an experimental environment and performed experiments for different frequencies, and through 2D localization using obtained sensor model, we meet with very confident results. The experimental results show a close correlation with the developed sensor model and the possibility of underwater localization systems using EM wave.

MOTIVATION: A new technique, mammalian green fluorescence protein (GFP) reconstitution across synaptic partners (mGRASP), enables mapping mammalian synaptic connectivity with light microscopy. To characterize the locations and distribution of synapses in complex neuronal networks visualized by mGRASP, it is essential to detect mGRASP fluorescence signals with high accuracy. RESULTS: We developed a fully automatic method for detecting mGRASP-labeled synapse puncta. By modeling each punctum as a Gaussian distribution, our method enables accurate detection even when puncta of varying size and shape partially overlap. The method consists of three stages: blob detection by global thresholding; blob separation by watershed; and punctum modeling by a variational Bayesian Gaussian mixture models. Extensive testing shows that the three-stage method improved detection accuracy markedly, and especially reduces under-segmentation. The method provides a goodness-of-fit score for each detected punctum, allowing efficient error detection. We applied this advantage to also develop an efficient interactive method for correcting errors. AVAILABILITY: The software is available on http://jinny.kist.re.kr.

Abstract Summary: We are constructing an ontology for traditional Korean medicine, and we started with medicinal materials to express the relationships between patients' symptoms, diseases and treatments. Biological materials and mineral resources have been used traditionally for patient treatments. The ontology includes various data related to these materials, such as their scientific names, parts of materials used, effectiveness and related oriental organ of the human body. Availability: http://tkm.kiom.re.kr/ontology/TraditionalKoreanMedicine.rdf-xml.owl (10–20 s using Internet Explorer.) Contact: [email protected] Supplementary information: Supplementary data are available at Bioinformatics online.

In this paper, we propose a new design of a flexible enveloping grasper for pick and place tasks with the low complexity in manipulation and task planning for the purpose of practical use in the near future. Flexible material for the grasper has many advantageous characteristics inherently including robustness against manipulation errors and the ability to increase contact area with a grasped object and the grasping force. Compliance of the grasper material also contributes to reduction in complexity of the processes such as the force control, sensor-motor coordination, and manipulation by self-adaptation. Two properties, flexibility and compliance, mentioned above help the proposed grasper minimize the internal forces in a passive manner and achieve the successful force distribution with self-adaptivity when performing enveloping grasping. In order to demonstrate our work, we have constructed 2 different prototypes of flexible enveloping grasper. Experimental results validate robust performances of the proposed grasper.

To improve the maneuverability of underwater vehicles, accurate thruster control is required based on a precise thrust model. Especially, in real systems, the ambient flow velocity and angle effects on the thrust force are very important, because vehicles are continuously moving and/or hovering with thruster actions. In this paper, we execute preliminary experiments with the model containing ambient flow effects as CAR (critical advance ratio) and CIA (critical incoming angle). The distinguished characteristics of the above model is the three axial flow states classification. The whole thrust map is divided into three states according to the state of ambient flow velocity, angle, and propeller shaft velocity. The several controllers are tested using the three flow states information. The preliminary experimental results show the best performance can be obtained by the open loop control with accurate model, because the thrust force cannot be measured directly, so the force map from the propeller shaft velocity to thrust force plays important roll in control performance. However, the DOB (disturbance observer) with closed loop controller is realistic solution, because it can compensate various unknown disturbances in real systems

The thruster is the crucial factor of an underwater vehicle system, because it is the lowest layer in the control loop of the system. However, an accurate and practical thruster model has not been utilized yet. In this paper, we propose an accurate and practical thrust modeling for underwater vehicles which considers the effects of ambient flow velocity. In this model, the axial flow velocity of the thruster, which is non-measurable, is represented by ambient flow velocity and propeller shaft velocity. Hence, contrary to previous models, the proposed model is practical since it uses only measurable states. Next, the whole thrust map is divided into three states according to the state of ambient flow and propeller shaft velocity, and one of the borders of the states is defined as Critical Advance Ratio (CAR). This classification explains the physical phenomenon of conventional experimental thrust maps. The proposed model is evaluated by comparing experimental data with numerical model simulation data, and it accurately covers overall flow conditions within ± 2N force error. The comparison results show that the new model’s matching performance is significantly better than conventional models’.

To improve the maneuverability of underwater vehicles, accurate thrust model is required. However, its modeling is not only difficult but also inaccurate, because it has highly nonlinear dynamics and many parameters which cannot be modeled exactly. Generally, a propeller's steady-state axial thrust is proportional to the signed square of propeller shaft velocity under bollard pull condition. And, it is reported that the velocity of the ambient flow of the thruster affects the thrust force largely. From these facts, we can easily make a reasonable inference that the thrust force may also be affected by the thruster spouting angle with respect to the vehicle moving direction. However, the effect of this non-parallel ambient flow is not well understood in the previous works. Hence, in this paper, a new parameter, what we call incoming angle, is introduced and its effect on thrust force is shown. Incoming angle is defined as the angle between the direction of thrust and that of vehicle velocity. The effect of this angle can be dominant during an underwater vehicle changes its direction, or an omni-directional vehicle carries out its task. The effect cannot be compensated by including the only parallel component of vehicle velocity into previous models. In this paper, based on the three axial flow states classification model, the effect of the incoming angle of ambient flow is analyzed, and critical incoming angle (CIA) is also defined to describe the thrust force states according to incoming angle. In order to characterize the effect of non-parallel ambient flow, the proposed model is evaluated by comparing experimental data with numerical model simulation data, and it accurately covers overall flow conditions within /spl plusmn/2N force error.

There are several kinds of singularity in controlling robotic manipulator. Among those, the kinematic and algorithmic singularity avoidance have been investigated intensively. What seems to be lacking, however, is excessive performance reduction in non-singular region and difficulty in performance tuning. In this paper, we develop a new method to solve kinematic and algorithmic singularities using task reconstruction approach. This new method drives the manipulator singularity-free path, and simultaneously guarantees task performance. And it can be easily extended singularity avoidance of multiple tasks case in task priority based method of redundant manipulators. The advantage and performance of the proposed method is validated by simulation works.

Underwater vehicle-manipulator (UVM) system becomes popular for various applications although it has complexity in control and analysis when compared to the vehicle itself. To resolve the complexity, firstly, thorough kinematic and dynamic analysis on the whole UVM system is performed. Then, the whole dynamics is divided into several meaningful terms. A robust coordinated motion control algorithm for autonomous UVM systems is considered next based on the analysis of divided dynamics. Since the vehicle has relatively smaller bandwidth than manipulators, active damping control with two-time scale approach for the manipulator is proposed. Using this algorithm, the manipulator controller can cover the vehicle state. The performance of the proposed algorithm is verified by computer simulations.

In this paper we describe an online solution for avoiding the occurrence of both algorithmic and kinematic singularities in task-priority based kinematic controllers of robotic manipulators. The proposed approach uses a secondary task correction and a successive task projection in order to maintain the measure of manipulability of the correspondent space augmentation approach over a minimum value. It shows a gain in performance and a better task error especially when working in proximity of singular configurations. It is particularly suitable for autonomous systems where an offline trajectory control scheme is often not applicable.

The dynamic stability of a mobile manipulator using ZMP compensation is considered. A unified approach for the two subsystems is formulated using a redundant scheme. First, to conserve the dynamic stability of the system, we define the performance index for the redundant system using the ZMP (zero moment point). This performance index represents the stability of the whole mobile manipulator system. Then, the redundancy resolution problem for optimizing the given performance index is solved using the null motion. Finally, the performance of this method is demonstrated by simulation study.

In autonomous system, it is important to establish a control scheme that works with stability even near singularity configurations. We describe an online trajectory control scheme that uses the manipulability measure as a distance criteria to avoid manipulator singularities. The proposed approach consists in a method for limiting the minimum value of the distance criteria. The performance is simply affected by the choice of the lower limit. Based on a real-time evaluation of the measure of manipulability, this method does not require a preliminary knowledge of the singular configurations. The proposed algorithm is validated by experimental results.

In this paper, we describe an online trajectory control scheme for robotic manipulators with dynamic change of the task priority. First, given tasks are reconstructed using a geometric projection on the given index function. Then, the reconstructed tasks are analyzed in the framework of task priority based method. From this analysis, the proposed algorithm is shown to have the property that the task priority is assigned dynamically. Using this algorithm, we easily set the criteria of changing task priority and estimate the performance of the given index function. Considering the measure of manipulability as a index function, the approach is suitable for avoiding kinematic singularities for autonomous operation of, for example, an underwater robot. The result shows a good performance near the singular configurations, as shown by simulation results.

We investigate the instantaneous impact phenomenon due to the collision between the manipulator and the environment for kinematically redundant manipulators. First, we define a normalized impact geometry in the viewpoint of normalized velocity change. We then reduce the impulsive effects under uncertain environment for redundant manipulators. A new impact performance index based on velocity direction is proposed. Finally, the performance of the proposed index is demonstrated by numerical examples and experiments.