Published papers

Due to the recent advances in Deep Neural Network (DNN) technologies, recognition and inference applications are expected to run on mobile embeddedsystems. Developing high-performance and power-efficient DNN engines becomesone of the important challenges for embedded systems. Since DNN algorithms orstructures are frequently updated, flexibility and performance scalability todeal with various types of networks are crucial requirement of the DNNaccelerator design. In this paper, we describe the architecture and LSI designof a flexible and scalable CNN accelerator called SNACC (Scalable NeuroAccelerator Core with Cubic integration) which consists of several processingcores, on-chip memory modules, and ThruChip Interface (TCI). We evaluate thescalability of SNACC with detailed simulation varying the number of cores andoff-chip memory access bandwidth. The results show that the energy efficiency of the accelerator becomes the highest in eight cores configuration with500MB/s off-chip bandwidth.

FiC (Flow-in-Cloud) is a multi-FPGA system to realize a monolithic large FPGA image with multiple cost-efficient mid-range FPGAs connected withfl exible and high bandwidth interconnection network. The FPGA is used as both for computational resource and the network circuit switch, and separating FPGA regions by partial reconfiguration (PR) design technique. In this paper, we will introduce the FPGA design methodology with the PR technique for the FiC system. Evaluation result shows that the proposed design methodology reduces design and configuration time, and it enables to replace the region for computation without stopping the network running on the same FPGA.

In this work, we propose a multi FPGA architecture with STDM network that aims to tackle compute-intensive applications such as neural networks training or pattern recognition in artificial intelligence while realizing high cost-performance and energy efficiency. To achieve this goal, optimizing different aspects of the system communication is a key challenge. In order to do this, a preliminary study on the application mapping for both the execution time and the number of slots for the STDM is carried out. An optimization based on a multi-criteria paradigm is implemented and the preliminary results show the possibility to optimize several parameters of the communication simultaneously alongside quantitative analyses of different architecture choices.

Network throughput has become an important issue for big-data analysis on Warehouse-Scale Computing (WSC) systems. It has been reported that randomly-connected inter-switch networks can enlarge the network throughput. For irregular networks, a multi-path routing method called k-shortest path routing is conventionally utilized. However, it cannot efficiently exploit longer-than-shortest paths that would be detour paths to avoid bottlenecks. In this work, a novel routing method called koptimized path routing to achieve high throughput is proposed for irregular networks. We introduce a heuristic to select detour paths that can avoid bottlenecks in the network to improve the average-case network throughput. Experimental results by network simulation show that the proposed k-optimized path routing can improve the saturation throughput by up to 18.2% compared to the conventional k-shortest path routing. Moreover, it can reduce the computation time required for optimization to 1/2760 at a minimum compared to our previously proposed method.

The techniques based on measuring the distance between objects and the camera itself have made progress in recent years. Through this technology, the system used for 3D imaging will provide a great convenience for people's lives. In this project, we focus on optimizing depth maps with better quality for the 3D scene display system on the car. During the 3D imaging process, to calculate the distance between a certain location and one of the specific points of the scene, it will produce two depth maps computed from two images as an intermediate product for 3D imaging processing. However, there is a great number of invalid pixels that distance cannot be measured. In dealing with such drawbacks, post-filtering has been introduced as a solution. In this project, we propose implementing the Weight Least Square (WLS) filter on FPGA, which can fill the invalid pixels by using the results of neighboring pixels. Through our approach, we can improve the quality of the depth map on M-KUBOS. Besides that, the optimized memory usage can successfully suit the size of the memory on-chip, and a 69.29% acceleration compared to the ARM core.

Body biasing can be used to control the leakage power and the performance of transistors by changing the threshold voltage after fabrication. Especially, a new process called Silicon-On-Thin BOX (SOTB) CMOS can control the balance of these two factors. When it is applied to a Coarse-Grained Reconfigurable Array (CGRA), the leakage power can be largely reduced by controlling precisely the bias with small domain size. On the other hand, the choices on bias voltages depend on the application executed on the platform, especially its mapping. In this paper, we propose to apply a multi-objective optimization for the application mapping and the body bias control on an energy efficient CGRA called CC-SOTB (Cool Mega Array Cube-SOTB). By using an NSGA-II algorithm for the mapping exploration and Integer Linear Programming (ILP) for the body bias control optimization, we show that it is possible to achieve better power consumption results than in previous works. For instance, in the case of a domain size of 2 rows by 1 column, it is possible to achieve a power reduction ratio up to 43.25%, compared to 21.09% previously, for the studied application. This is however achieved at the cost of a bigger mapping width. Nonetheless, the exploration allows to have finer analyses about both mapping and consumption. Indeed, these promising results show that optimizing the application mapping simultaneously with the body bias control can provide more interesting results, giving deeper quantitative information about trade-off possibilities.

Low-power embedded systems often require simple but flexible functionality for data management. Key-value store is one of data stores, which provides simple API and is easy to scale-out. We developed a dedicated key-value store core as a low-power embedded storage. In this paper, key-value store chip is fabricated with Renesas SOTB 65-nm process technology. We evalute the real chip in terms of power consumption and operating frequency by tuning VDD and body bias voltage. The chip achieved 11.2mW as power consumption with 0.7V VDD if the target frequency is 40MHz. When data rate is low, the system can reduce power consumption by tuning VDD and body bias voltage.

As the scale of supercomputers has increased, the end-to-end network latency has become non-negligible on the performance of parallel applications. For this reason, low-latency network topologies achieving the smallest number of hops among switches are drawing attention. The Order/Degree Problem to find the topology with the smallest diameter has revealed that a topology based on graph symmetry and simulated annealing (SA) is optimal for a wide range of system sizes. Focusing on the characteristics of clustering and symmetry of this graph, we extend the method to guarantee the lower limit of the number of cables in each cabinet. The total cable length and cost can be minimized as well as minimizing the deterioration of communication delay. Compared with the conventional network topologies, the proposed network topology can attain higher performance with the smaller total cable length and cost.

This paper proposes a level-shifter free (LSF) approach for multi-VDD design to employ a combination of body bias control and a superior threshold-voltage (Vt) modulation capability of SOTB (Silicon on Thin BOX) devices. We applied this approach to a microprocessor test chip with low-voltage (VDDL) and high-voltage (VDDH) domains, and fabricated it in a 65nm SOTB technology. Measurement results demonstrated that the chip correctly operates at VDDL=0.6V and VDDH=1.2V under the reverse-body-bias (RBB) of 2V for pMOS transistors in the VDDH domain while suppressing the static dc current.

Level shifters to convert signal swings from low-voltage (VDDL) to high-voltage (VDDH) are required at the boundary of voltage domains in SoC employing multiple supply voltages. However, they cost delay, power and area in addition to increasing the complexity of physical design. This paper proposes a level-shifter-less (LSL) approach to use a reverse body bias (RBB) at pMOS transistors in the VDDH domain and superior threshold-voltage modulation capability of FD-SOI devices. Simulation results and measurements of a fabricated chip showed that the chip applying the LSL approach correctly operates at VDDL = 0.6 V and VDDH = 1.2 V under 2 V pMOS RBB while suppressing the static dc current in the VDDH domain. We also demonstrate that adaptive RBB control for pMOS can maintain effectiveness of this approach under process and temperature variations.

Interconnection network ideally transfers the maximum amount of communication dataset within the least amount of time to fully exploit the parallelism of target applications on parallel computer systems. To this goal, we propose a selective data-compression interconnection network. Data compression virtually increases the effective network bandwidth, while each compute node introduces additional latency overhead to perform (de-)compression operation to end-To-end communication latency. To minimize the effect of the compression latency overhead on the end-To-end communication latency, we selectively apply a compression technique to a packet. The compression operation is taken for long packets and is also taken when network congestion is detected at a network interface. Evaluation results show that simple lossless and lossy compression algorithms have up to 3.0 and 1.8 compression ratios for integer and floating-point communication data in some parallel applications, respectively, while the lossy compression algorithm successfully satisfies the required quality of results. Through a cycle-network simulation, the selective compression method using the above compression algorithms improves by up to 46% the network throughput with the moderate increase of the communication latency of short packets.

Three-dimensional stack memory that provides both high-bandwidth access and large capacity is a promising technology for next-generation computer systems. While a large number of memory cubes increase the aggregate memory capacity, the communication latency and power consumption would become significant due to its low-radix large-diameter packet network. In this context, we propose a memory-cube network called Diagonal Memory Network (DMN). A diagonal network topology, its floor layout, and its lightweight router are designed for low-latency and low-voltage memory-read communication. Our evaluation results show that a DMN router decreases 31% of the hardware resources than a conventional virtual-channel router. The DMN router reduces 13% and 67% energy consumption to transit a packet along with the original datapath and bypassing datapath, respectively.

The M-KUBOS/PYNQ cluster is proposed as an energy efficient computing server for MEC(Multi-Access Edge Computing) which treats various tasks including time-critical jobs at 5G base stations. M-KUBOS boards each of which provides the largest model of Zynq UltraScale+ are connected with high speed serial GTH links and STDM (static time division multiplexing) switches. PYNQ(Python productivity for Zynq), an open-source software platform by Xilinx was introduced so that the multi-Zynq system can be used as a multitenant server. The PYNQ cluster can be programmed by extension of a single PYNQ system, and RESTful APIs are introduced for runtime management of multiple Zynq boards from a remote server. The design examples and evaluation results demonstrate that the M-KUBOS/PYNQ cluster achieved 1.0-17.2 times the performance and 1.75-85.8 times the performance per watt of the recent Ryzen-7 server.

Coarse-grained reconfigurable architectures (CGRAs) provide high energy efficiency with word-level programmability rather than bit-level ones such as FPGAs. The coarser reconfigurability brings about higher energy efficiency and reduces the complexity of compiler tasks compared to the FPGAs. However, application mapping process for CGRAs is still time-consuming. When the compiler tries to map a large and complicated application data-flow-graph(DFG) onto the reconfigurable fabric, it tends to result in inefficient resource use or to fail in mapping. In case of failure, the compiler must divide it into several sub-DFGs and goes back to the same flow. In this work, we propose a novel partitioning method based on a genetic algorithm to eliminate the unmappable DFGs and improve the mapping quality. In order not to generate unmappable sub-DFGs, we also propose an estimation model which predicts the mappability and resource requirements using a DGCNN (Deep Graph Convolutional Neural Network). The genetic algorithm with this model can seek the most resource-efficient mapping without the back-end mapping process. Our model can predict the mappability with more than 98% accuracy and resource usage with a negligible error for two studied CGRAs. Besides, the proposed partitioning method demonstrates 53-75% of memory saving, 1.28-1.39x higher throughput, and better mapping quality over three comparative approaches.

FiC (Flow-in-Cloud) is a multi-FPGA system to achieve monolithic large scale FPGA with across multiple FPGAs for computational-heavy applications. The FiC system employs multiple economical mid-range FPGAs and connect them directly with flexible and high bandwidth interconnection network. We developed FiCSW board as a 1st generation of FPGA board in the FiC project. The FiCSW design, since the FPGA is used as both computational resource and the network circuit switch, there are no host server for FPGA needed like other server-based multi FPGA architecture. This design strategy helps scalable FPGA fabric in low-cost. However, from the system manageability's perspective, the strategy is hard to handle multiple FPGA nodes without management system. In this paper, we mainly focused on the management system on the FiCSW multi-FPGA system, introducing the system architecture and implementation, and working application on the FiCSW prototype system.

Non-volatile power gating introducing MTJ into a flip-flop is an attractive technique for IoT devices. However, the time required for a successful store operation to MTJs varies at non-volatile flip-flops (NVFFs) due to process and voltage variation. This paper describes implementation and real-chip evaluation of a non-volatile Coarse Grained Reconfigurable Array employing verify-and-retryable NVFF (VR-NVFF) circuits and two-step store (TSS) control by instructions of a microcontroller. Test chips fabricated in a 40nm CMOS/MTJ hybrid technology have demonstrated that the TSS approach using both 30ns/100ns store operations reduces power by 37-42% when running an image processing program at 100MHz.

Variable Pipeline Cool Mega Array (VPCMA) is a low power Coarse Grained Reconfigurable Architecture (CGRA) based on the concept of CMA (Cool Mega Array). It provides a pipeline structure in the PE array that can be configured so as to fit target algorithms and required performance. Also, VPCMA uses the Silicon On Thin Buried oxide (SOTB) technology, a type of Fully Depleted Silicon On Insulator (FDSOI), so it is possible to control its body bias voltage to provide a balance between performance and leakage power. In this paper, we study the optimization of the VPCMA body bias while considering simultaneously its variable pipeline structure. Through evaluations, we can observe that it is possible to achieve an average reduction of energy consumption, for the studied applications, of 17.75% and 10.49% when compared to respectively the zero bias (without body bias control) and the uniform (control of the whole PE array) cases, while respecting performance constraints. Besides, it is observed that, with appropriate body bias control, it is possible to extend the possible performance, hence enabling broader trade-off analyzes between consumption and performance. Considering the dynamic power as well as the static power, more appropriate pipeline structure and body bias voltage can be obtained. In addition, when the control of VDD is integrated, higher performance can be achieved with a steady increase of the power. These promising results show that applying an adequate optimization technique for the body bias control while simultaneously considering pipeline structures can not only enable further power reduction than previous methods, but also allow more trade-off analysis possibilities.

We developed a PYNQ cluster called M-KUBOS that consists of economical Zynq boards that are interconnected through low-cost high-performance GTH serial links. For the software environment, we employed the PYNQ open-source software platform. The PYNQ cluster is anticipated to be a multi-access edge computing (MEC) server for 5G mobile networks. We implemented the ResNet-50 inference accelerator on the PYNQ cluster for image recognition of MEC applications. By estimating the execution time of each ResNet-50 layer, layers of ResNet-50 were divided into four boards so that the execution time of each board would be as equal as possible for efficient pipeline processing. Owing to the PYNQ cluster in which FPGAs were directly connected by high-speed serial links, stream processing without network bottlenecks and pipeline processing between boards were readily realized. The implementation achieved 292 GOPS performance, 75.1 FPS throughput, and 5.15 GOPS/W power efficiency. It achieved 17 times faster speed and 86 times more power efficiency compared to the implementation on the CPU, and 3.8 times more power efficiency compared to the implementation on the GPU.

The next generation high-performance computing platforms will need to support exascale computing. A promising path in achieving exascale is to embrace heterogeneity and specialised computing in the form of reconfigurable accelerators. However, assessing the feasibility of heterogeneous exascale systems requires fast and accurate performance prediction. This paper proposes PERKS, a novel performance estimation frame-work for reconfigurable dataflow platforms (RDPs). PERKS uses machine and application parameters to build an analytical model for predicting the performance of multi-Accelerator systems. Moreover, model calibration is automatic, making the model flexible and usable for different machine configurations and applications. Our experimental results demonstrate that PERKS can predict the performance of current workloads and RDPs with an accuracy above 95%. We also demonstrate how the modelling scales to exascale workloads and exascale platforms.

This paper presents an approach for analysing, modelling and predicting application performance of large-scale heterogeneous platforms. Our approach combines analytical and statistical modelling techniques, and aims to: (1) identify and characterise code regions that are the most promising candidates to benefit from acceleration; (2) provide statistical models that predict application behaviour for unobserved inputs; and (3) predict performance gain with different system architectures.