Published papers

Although low power consumption is a significant advantage of FPGA clusters, almost no power analyses with real systems have been reported. This study reports the detailed power consumption analyses of two FPGA clusters, namely, M-KUBOS and FiC, with power measurement tools and real applications. In both clusters, the type of logic design shells determines the base power consumption. For building clusters, the power for node communication links is mainly determined by the number of activated links and not influenced by the number of actually used links. Therefore, applying the link aggregation technique does not affect the power consumption. Increasing the clock frequency of the application logic mildly increases the power consumption. The obtained results suggest that the best way to reduce the total power consumption of an FPGA cluster and improve its performance is to use the minimum number of links for the application, apply link aggregation, and aggressively increase the clock frequency.

There are emerging services for the transports and nursing with multiple robots has become more familiar to our society. Considering the increasing demand for automatic multiple robotic services, it appears the research into automatic multiple robotic services is not satisfactory. Specifically, the issues of power consumption of these robots, and its potential reduction have not been sufficiently discussed. In this research, we propose a method and system to reduce the aggregated power consumption of multiple robots by modelling the characteristics of the hardware and service of each robot. We firstly discuss the prediction model of the robot and improve the formula with consideration of its use in a wide range of situations. Then, we achieve the objective of reducing the aggregate power consumption of multiple robots, using consumption logs and re-allocating tasks of them based on the power consumption prediction model of the individual robot. We propose the design and develop a system using ROS (Robot Operating System) asynchronous server to collect the data from the robots, and make the prediction model for each robot, and reallocate tasks based on the findings of the optimized combination on the server. Through the evaluation of the design and implementation with the proposed system and the actual robot Zoom (GR-PEACH + Rasberry pi), we achieve an average power reduction effect of 14%. In addition, by offloading high-load processing to an edge server configured with FPGA instead the Intel Core i7 performance computer, we achieved and increase in processing speed of up to about 70 times.

In parallel processing applications, a few worker nodes called "stragglers", which execute their tasks significantly slower than other tasks, increase the execution time of the job. In this paper, we propose a network switch based straggler handling system to mitigate the burden of the compute nodes. We also propose how to offload detecting stragglers and computing their results in the network switch with no additional communications between worker nodes. We introduce some approximate techniques for the proxy computation and response at the switch; thus our switch is called "ApproxSW." As a result of a simulation experiment, the proposed approximation based on task similarity achieves the best accuracy in terms of quality of generated Map outputs. We also analyze how to suppress unnecessary proxy computation by the ApproxSW. We implement ApproxSW on NetFPGA-SUME board that has four 10Gbit Ethernet (10GbE) interfaces and a Virtex-7 FPGA. Experimental results shows that the ApproxSW functions do not degrade the original 10GbE switch performance.

Even though many optimization methods for CGRAs (Coarse-Grained Reconfigurable Architectures) have been proposed, aggressive power optimization still remains a complex problem to be solved. Moreover, the developments of these methods have mainly been proven on the basis of simulations. Therefore, the questions remains whether they can be applied for a real chip. Here, we consider a real implemented low power CGRA called CCSOTB2, and explore the possibility of the power reduction for this design. This paper proposes to use a metaheuristic method to optimize the power while considering all configurable factors of the CGRA, especially the mapping of an application. This methodology can generate mappings with their related pipeline structure and body bias control automatically. Optimized configurations to use on the real chip are obtained with this methodology and allow to measure the power consumption. The experimental results show a power reduction of 14.2% in average, when compared to a previously-used mapping method which cannot consider body bias and pipeline structure. In addition, the proposed method enables users to select a mapping from various solutions depending on performance requirement and trade-off possibilities (e.g. throughput vs power consumption).

The building block computing system can build various systems by connecting small-sized chips with inductive coupling wireless chip-to-chip connection TCI (Through Chip Interface). First, we have developed TCI IP by using the Renesas 65nm SOTB process and verified it with a simple TEG chip. Then several family chips have been developed with the IP; Geyser-TT (MIPS R3000 compatible processor), SNACC (neural network accelerator), CC-SOTB2 (improved CGRA), KVS (the accelerator for Key-Value-Store database), and SMTT (Shared Memory for Twin Tower). All of these chips worked alone without problems. However, when these chips were stacked to construct a system, problems were found on some combination of chips that TCI did not operate as designed. To investigate the cause of the trouble, we have developed a sophisticated tester chip called the TCI Tester. It provides three modes: RAW mode in which handshake lines can be directly controlled, CUBE mode in which the TCI Tester can write the data into the stacked target chips, and LOOP mode which allows continuous write and read operation for a long time working test. Through the evaluation by using a two-TCI Tester stacking system, the following appeared. (1) The maximum transfer frequency of the TCI IP is much lower than 50MHz which is the target value. Only 14MHz for the downward and 9MHz for the upward. (2) The performance of the upward transfer is lower than that of the downward. The poor upward data transfer performance comes from that the voltage drop on the power grid becomes large because of the location of the upward transmitter coil. We can improve it by increasing the number of power pads for the transmitter and enhancing the power grid. However, with the low transfer frequency, the TCI IP worked more than a day continuously.

Recent IoT devices require extremely low standby power consumption, while a certain performance is needed during the active time, and Coarse-Grained Reconfigurable Arrays (CGRAs) have received attention because of their high energy efficiency. For further reduction of the standby energy consumption of CGRAs, the leakage power for their configuration memory must be reduced. Although the power gating is a common technique, the lost data in flip-flops and memory must be retrieved after the wake-up. Recovering everything requires numerous state transitions and considerable overhead both on its execution time and energy. To address the problem, Non-volatile Cool Mega Array (NVCMA), a CGRA providing non-volatile flip-flops (NVFFs) with spin transfer torque type non-volatile memory (NVM) technology has been developed. However, in general, non-volatile memory technologies have problems with reliability. Some NVFFs are stacked-at-0/1, and cannot store the data in a certain possibility. To improve the chip yield, we propose a mapping algorithm to avoid faulty processing elements of the CGRA caused by the erroneous configuration data. Next, we also propose a method to add an error-correcting code (ECC) mechanism to NVFFs for the configuration and constant memory. The proposed method was applied to NVCMA to evaluate the availability rate and reduction of write time. By using both methods, the average availability ratio of 94.2% was achieved, while the average availability ratio of the nine applications was 0.056% when the probability of failure of the FF was 0.01. The energy for storing data becomes about 2.3 times because of the hardware overhead of ECC but the proposed method can save 8.6% of the writing power on average.

Coarse-grained reconfigurable architectures (CGRAs) is one of the suitable devices for IoT (Internet of Things) and edge-computing because of their high energy efficiency and programmability. The CGRAs process a compute-intensive part of an application program (especially a loop part) more efficiently than general purpose processors. CMA (Cool Mega Array) is an energy-conscious CGRA with a task-level reconfiguration instead of a cycle-level one. However, the CMA faces some limitations related to data management because of the aggressive pursuit of power saving. In this paper, we introduce a new CMA architecture VPCMA2 to relax the constraints and to improve energy efficiency. Then, we implement it with a 65-nm process technology to evaluate a hardware overhead due to the improvement. According to the evaluation results, the new design does not influence its maximum operating frequency. Although new functionalities brought about 17% power overhead and 10% area overhead, a remarkable improvement of application mappability and data handling was achieved.

Multi-FPGA systems have been receiving a lot of attention as a low cost and energy efficient system for Multi-access Edge Computing (MEC). For such purpose, a bare-metal multi-FPGA system called FiC (Flow-in-Cloud) is under development. In this paper, we introduce the FiC multi FPGA cluster which is applied partial reconfiguration (PR) FPGA design flow to support online user defined accelerator replacement while executing FPGA interconnection network and its low-level multiple FPGA management software called remote PR manager. With the remote PR manager, the user can define the FiC FPGA cluster setup by JSON and control the cluster from user application with the cooperation of simple cluster management tool / library called ficmgr on the client host and REST API service provider called ficwww on Raspberry Pi 3 (RPi3) on each node. According to the evaluation results with a prototype FiC FPGA cluster system with 12 nodes, using with online application replacement by PR and on-the-fly FPGA bitstream compression, the time for FPGA bitstream distribution was reduced to 1/17 and the total cluster setup time was reduced by 21∼57% than compared to cluster setup with full configuration FPGA bitstream.

The Boolean satisfiability problem (SAT) is an NP-complete combinatorial optimization problem, where fast SAT solvers are useful for various smart society applications. Since these edge-oriented applications require time-critical control, a high speed SAT solver on FPGA is a promising approach. Here the authors propose a novel FPGA implementation of a bio-inspired stochastic local search algorithm called 'AmoebaSAT' on a Zynq board. Previous studies on FPGA-AmoebaSATs tackled relatively smaller-sized 3-SAT instances with a few hundred variables and found the solutions in several milli seconds. These implementations, however, adopted an instance-specific approach, which requires synthesis of FPGA configuration every time when the targeted instance is altered. In this paper, a slimmed version of AmoebaSAT named 'AmoebaSATslim,' which omits the most resource-consuming part of interactions among variables, is proposed. The FPGA-AmoebaSATslim enables to tackle significantly larger-sized 3-SAT instances, accepting 30,000 variables with 130, 800 clauses. It achieves up to approximately 24 times faster execution speed than the software-AmoebaSATslim implemented on a CPU of the x86 server.

Stereo depth estimation has become an attractive topic in the computer vision field. Although various algorithms strive to optimize the speed and the precision of estimation, the energy cost of a system is also an essential metric for an embedded system. Among these various algorithms, Semi-Global Matching (SGM) has been a popular choice for some real-world applications because of its accuracy-and-speed balance. However, its power consumption makes it difficult to be applied to an embedded system. Thus, we propose a robust stereo matching system, RT-libSGM, working on the Xilinx Field-programmable gate array (FPGA) platforms. The dedicated design of each module optimizes the speed of the entire system while ensuring the flexibility of the system structure. Through an evaluation running on a Zynq FPGA board called M-KUBOS, RT-libSGM achieves state-of-the-art performance with lower power consumption. Compared with the original design (libSGM), when working on the Tegra X2 GPU, RT-libSGM runs 2 × faster at a lower energy cost.

Due to the recent advances in Deep Neural Network (DNN) technologies, recognition and inference applications are expected to run on mobile embedded systems. Developing high-performance and power-efficient DNN engines becomes one of the important challenges for embedded systems. Since DNN algorithms or structures are frequently updated, flexibility and performance scalability to deal with various types of networks are crucial requirement of the DNN accelerator design. In this paper, we describe the architecture and LSI design of a flexible and scalable CNN accelerator called SNACC (Scalable Neuro Accelerator Core with Cubic integration) which consists of several processing cores, on-chip memory modules, and ThruChip Interface (TCI).

Wireless interconnects based on inductive coupling technology are compelling propositions for designing 3-D integrated chips. This work addresses the heat dissipation problem on such systems. Although effective cooling technologies have been proposed for systems designed based on Through Silicon Via (TSV), their application to systems that use inductive coupling is problematic because of increased wireless-communication distance. For this reason, we propose two methods for designing sparse 3-D chips layouts and Networks on Chip (NoCs) based on inductive coupling. The first method computes an optimized 3-D chip layout and then generates a randomized network topology for this layout. The second method uses a standard stack chip layout with a standard network topology as a starting point, and then deterministically transforms it into either a "staircase" or a "checkerboard" layout. We quantitatively compare the designs produced by these two methods in terms of network and application performance. Our main finding is that the first method produces designs that ultimately lead to higher parallel application performance, as demonstrated for nine OpenMP applications in the NAS Parallel Benchmarks.

Superpixel segmentation is a very popular image segmentation technique used in various computer vision tasks. Recently, a number of superpixel algorithms have been proposed in literature. One such algorithm is considered as the-state-of-the-art in superpixel segmentation: Simple Linear Iterative Clustering or SLIC. However, its original implementation has a long execution time on high performance processors designed within the common mobile and enterprise applications, as well on high-end processors such as Intel Xeon. Overall, the execution time for single-threaded implementation is considered critical for real-time or near real-time applications. In this paper, we explore the possibility of accelerating parts of the SLIC image segmentation critical for performance, by designing the image segmentation accelerator for Intel's Arria 10 SoC. We propose a novel architecture to enable hardware acceleration by addressing the problem of hardware/software partitioning to minimize the overall program latency.

NISQ (Noisy, Intermediate-Scale Quantum) computing requires error mitigation to achieve meaningful computation. Our compilation tool development focuses on the fact that the error rates of individual qubits are not equal, with a goal of maximizing the success probability of real-world subroutines such as an adder circuit. We begin by establishing a metric for choosing among possible paths and circuit alternatives for executing gates between variables placed far apart within the processor, and test our approach on two IBM 20-qubit systems named Tokyo and Poughkeepsie. We find that a single-number metric describing the fidelity of individual gates is a useful but imperfect guide. Our compiler uses this subsystem and maps complete circuits onto the machine using a beam search-based heuristic that will scale as processor and program sizes grow. To evaluate the whole compilation process, we compiled and executed adder circuits, then calculated the Kullback-Leibler divergence (KL-divergence, a measure of the distance between two probability distributions). For a circuit within the capabilities of the hardware, our compilation increases estimated success probability and reduces KL-divergence relative to an error-oblivious placement.

Reinforcement learning agents have shown very good results in robot control and navigation tasks, allowing robots to learn how to interact with an environment appropriately in a model-free manner. However, real-world robot systems have strict latency, power, and cost constraints, thus requiring special hardware consideration for the demanding computations of neural networks. Furthermore, reinforcement learning networks should be able to interface efficiently with the various other robot components. To address these challenges, we propose a method for applying FPGA hardware accelerators to robotics reinforcement learning agents at the inference stage and seamlessly integrating the FPGA hardware module to the robot system by automatically wrapping it in a Robot Operating System 2 (ROS2) node. The proposed system is evaluated in three OpenAI gym control environments: Cartpole-v1, Acrobot-v1, and Pendulum-v0. In the evaluation, both quantized and non-quantized reinforcement learning neural networks are used, and the proposed FPGA system is observed to provide up to a 3.69x speed up and up to 52.7x better performance per watt when compared to an agent running on a ROS2 node on a modern CPU.

We propose an autonomous vehicle controlled by FPGAs. In our design, considering embedded systems, we apply the binarized neural networks (BNNs) which can realize a satis-fying result in high speed and accuracy to recognize pedestrians and some obstacles on a given road. To detect the traffic light, a passive camera-based pipeline is applied. Furthermore, the implementation of road lane detection is based on color selection algorithm, Canny Edge Detection, and Hough Transformation. The proposed design is realized by two Xilinx boards: PYNQ-Z1 and Zynq-Xc7Z010. These two FPGA boards cooperate with each other through a shared network cable. In the proposed design, the resource used by Zynq-Xc7Z010 can be greatly reduced and the inference time on the FPGA has been thousands times faster than the software implementation.

Convolutional Neural Networks (CNN) have been widely used for various computer vision tasks. While GPUs are the most common platform for CNN implementation, FPGAs are promising alternatives to provide better energy efficiency. Recent work demonstrates the potential of network quantization to reduce the model size and enhance computation efficiency while maintaining comparable accuracy to the full precision counterparts. Quantized CNN is especially suitable for FPGA implementation due to the presence of values with non-trivial bitwidth. In this paper, we present the design of an FPGA-based accelerator for losslessly quantized CNNs using High Level Synthesis tool. The experiment result shows that our design achieves 12.9 GOPS/Watt for quantized Alexnet on Imagnet Dataset.

This paper presents the implementation of lane line detection on FPGA and Python. Lane line detection consists of three functions, median blur, adaptive threshold, and Hough transform. We implemented only accumulation of Hough transform on FPGA. Although the Hough transform cannot be implemented on a low-end FPGA board if implemented directly, by reducing ρθ space, it was successfully implemented on a low-end FPGA board. The rest of the Hough transform was implemented using Python's NumPy and SciPy, and OpenCV. Although it was very easy to write, it did not become a bottleneck for the whole process because of its effectiveness. As a result, we could achieve a 3.9x speedup compared to OpenCV and kept the developing cost down. When implementing median blur and adaptive threshold on an FPGA, we could achieve a 6.34x speedup.

Coarse-grained reconfigurable architectures (CGRAs) are expected to be used for embedded systems, Internet of Things (IoT) devices, and edge computing thanks to their high-energy efficiency and programmability. In essence, a CGRA is an array of numerous processing elements. To exploit this abundant computation resource, a compiler for CGRAs has to fulfill more tasks compared that for general-purpose processors. Therefore, many studies have proposed optimization methods, especially for application mapping, because the performance and energy efficiency strongly depend on optimization at compile time. However, many works focus only on performance improvement or resource minimization, although such optimization objectives are not always appropriate when considering various use cases. In this work, we propose GenMap, an application mapping framework using multiobjective optimization based on a genetic algorithm so that users can set optimization criteria as needed. Besides, it provides aggressive power optimization using our dynamic power model and leakage minimization technique. The proposed method is applied to three fabricated CGRA chips for evaluation. Experimental results show that GenMap achieves 15.7% reduction of wire length while keeping processing element utilization when compared with conventional methods. In addition, according to real chip experiments, 12.1%-46.8% of energy consumption is reduced, and up to 2\times speedup is archived for several architectures when compared with other two approaches.