1,721,317 research outputs found
Validating Full-System RISC-V Simulator: A Systematic Approach
No full textRISC-V-based Systems-on-Chip (SoCs) are witnessing a steady rise in adoption in both industry and academia. However, the limited support for Linux-capable Full System-level simulators hampers development of the RISC-V ecosystem. We address this by validating a full system-level simulator, gXR5 (gem5-eXtensions for RISC-V), against the SiFive HiFive Unleashed SoC, to ensure performance statistics are representative of actual hardware. This work also enriches existing methodologies to validate the gXR5 simulator against hardware by proposing a systematic component-level calibration approach. The simulator error for selected SPEC CPU2017 applications reduces from 44% to 24%, just by calibrating the CPU. We show that this systematic component-level calibration approach is accurate, fast (in terms of simulation time), and generic enough to drive future validation efforts.ES
3D-ICE 3.0: efficient nonlinear MPSoC thermal simulation with pluggable heat sink models
Full text linkThe increasing power density in modern high-performance multi-processor system-on-chip (MPSoC) is fueling a revolution in thermal management. On the one hand, thermal phenomena are becoming a critical concern, making accurate and efficient simulation a necessity. On the other hand, a variety of physically heterogeneous solutions are coming into play: liquid, evaporative, thermoelectric cooling, and more. A new generation of simulators, with unprecedented flexibility, is thus required. In this paper, we present 3D-ICE 3.0, the first thermal simulator to allow for accurate nonlinear descriptions of complex and physically heterogeneous heat dissipation systems, while preserving the efficiency of latest compact modeling frameworks at the silicon die level. 3D-ICE 3.0 allows designers to extend the thermal simulator with new heat sink models while simplifying the time-consuming step of model validation. Support for nonlinear dynamic models is included, for instance to accurately represent variable coolant flows. Our results present validated models of a commercial water heat sink, and an air heat sink plus fan that achieve an average error below 1 degree Celsius and simulate, respectively, up to 3x and 12x faster than the real physical phenomena.ES
Multi-level Optimization Methodologies for Thermally Reliable Multi-Core Architectures
No full textES
Thermal Simulation of Liquid-Cooled Integrated Circuits
No full textGrowing demands for increased functionality in consumer electronics and for information technology-enabled services in various sectors have resulted in the rise of high-performance multiprocessor system-on-chips (MPSoCs) and three-dimensionally stacked integrated circuits (3D ICs), that are deployed in various electronic devices and data centers to cope with these demands. This, in turn, has resulted in an alarming rise in electronic heat dissipation that now matches the levels typically encountered in nuclear reactors. On a small scale, the increased heat flux in ICs undermines the thermal reliability and lifetimes of these devices. On a large scale, it dramatically increases the cooling costs and the corresponding energy expenditure in data centers, thus escalating their carbon footprint globally to equal that of the airline industry, according to recent reports. Conventional copper-based air-cooled heat sinks and thermal packages are increasingly falling short in addressing these problems. Hence, advanced thermal packaging based on single- and two-phase liquid cooling of electronics have been recently proposed to cool integrated circuits to safe operating temperatures in a cost-effective manner. While possessing desirable properties compared to conventional heat sinks such as increased cooling capacity and energy-efficiency, liquid cooling of ICs does present various design challenges such as cost of technology migration, the design of micro-scale channels on silicon dies, implementation of electrical isolation and safe delivery of the coolant to the thermal package, optimization of electronic performance vis a vis cooling energy expenditure and structural reliability of the circuits. Electronic design automation (EDA) tools are needed during early stages of the design to evaluate these aspects, explore vast design spaces, develop control and management policies, reduce time-to-market, and minimize manufacturing and operating costs. At the heart of these tools is a fast compact thermal modeling method that can simulate the temperatures and cooling energies accurately for liquid-cooled ICs, in order to enable quick design space explorations at an early stage. In this thesis, compact thermal models for 2D/3D ICs are proposed to address this need. A compact thermal model for ICs with single-phase liquid cooling called 3D-ICE, a semi-analytical thermal model for the optimized design of liquid-cooled ICs and a compact thermal model for ICs with two-phase cooling called STEAM are proposed. The accuracy of these thermal models are validated against temperature measurements from real IC test stacks. In addition, these models are demonstrated to be orders of magnitude faster than conventional fine-grained simulators. Finally, two methods for accelerating thermal simulation are proposed to enhance the applicability of these methods in the early-stage design of 2D/3D ICs and MPSoCs with liquid cooling and other advanced thermal packages.ES
Multi-Objective Management of Multiprocessor Systems: From Heuristics to Reinforcement Learning
No full textIn my thesis, I reveal several already-existing and emerging challenges in multi-objective management of multiprocessor systems, and address them through novel solutions, from heuristics to RL, depending on the complexity of the problem. Conventional multi-objective management of multiprocessor systems mostly focuses on hot spots as the main factor of lifetime reliability. For modern multiprocessor systems and workloads, thermal stress has become the dominant factor in determining the Mean Time-To-Failure (MTTF). Together with the advances in multiprocessor systems, cooling technologies have been also progressively improving. As a result, existing Dynamic Thermal Management (DTM) policies should adapt to these emerging challenges and technologies to further improve the lifetime reliability. Finally,
Therefore, I first propose a holistic, yet fast thermal stress-aware heuristic approach. The results demonstrate that the lifetime reliability can increase by up to 47%.
Then, I show how emerging cooling technologies, such as two-phase liquid-cooling thermosyphon, necessitates adapting conventional heuristics to gain the greatest possible advantage from all its potential. My proposed approach decreases thermal hot spots and thermal stress by up to 10 oC and 45%, respectively, with 45% less cooling power consumption.
Input-dependent workload variation in emerging applications and services, such as multimedia streaming makes power and performance management more challenging. Thus, I propose a machine learning-based framework for workload prediction and throughput estimation using hardware events available on modern multiprocessor systems. The proposed machine learning framework achieves 3.4x higher throughput with 15% less power consumption for High Efficiency Video Coding (HEVC), as a test-case application.
I address runtime management and design space search of large and dynamic environments through RL. In particular, I first propose an RL-based framework to enable proactive fan speed control along with DVFS and workload allocation, providing up to 40% cooling power savings without any thermal constraint violations.
Second, I address multimedia workload allocation of HEVC encoder on heterogeneous Systems-on-Chip (SoCs) through RL, achieving 20% higher compared to state of the arts.
Third, I propose an RL-based approach that enables joint optimization of application- and system-level parameters, improving power consumption, performance, and average temperature of multiprocessor systems by 13%, 15%, and 10%, respectively, while improving the video quality and video compression of HEVC encoders, as a use-case application, by up to 1.19 dB and 24%.
Then, to speed up design space search, I propose a Multi-Agent Reinforcement Learning (MARL) approach for multi-objective runtime management of multiprocessor systems. I use HEVC encoder as a test-case application, where MARL can enhance QoS violations by 5x, while speeding up the learning phase 15x.
Finally, I address hyperparameter optimization of Convolutional Neural Network through a novel MARL approach. My proposed solution can reduce the model size, training time, and inference time by up to, respectively, 83x, 52\%, and 54\% without any degradation in accuracy.ES
Distributed Machine Learning Targeting Embedded Systems for Epilepsy Detection
No full textEpileptic seizure detection and monitoring is critical in healthcare, particularly for individuals requiring continuous oversight. Current methods, primarily based on electroencephalogram (EEG) technologies, face limitations due to their complexity in terms of acquisition, processing, comfort, and ease of use by the patient. Innovations in signal processing and machine learning have facilitated advances beyond traditional monitoring methods. In this context, an innovative approach that uses electrocardiogram (ECG) signals for seizure detection is introduced, offering a viable alternative to the complexities associated with EEG methodologies. This novel approach not only simplifies the monitoring process, but also significantly improves user comfort by avoiding the invasive nature of conventional EEG techniques.
In addition, the Multi-to-Single Knowledge Distillation (M2SKD) framework has been developed, aimed specifically at optimizing the balance between computational efficiency and diagnostic precision in wearable devices. This transition from a multi-biosignal to a single-biosignal model effectively mitigates the critical trade-offs between power consumption and algorithmic performance, crucial for the functionality of wearable technologies. This approach ensures that wearable systems maintain high accuracy levels, a fact substantiated by extensive simulations evaluating the framework's performance across various edge computing platforms.
Addressing the prevalent challenges of high memory and computational demands in neural network training within resource-constrained environments, the novel Bio-BPfree methodology is introduced. This innovative approach abandons the conventional backpropagation technique, which is not suitable for low-power environments, in favor of a specialized learning process tailored for wearable biomedical systems. By deploying a unique set of objective functions and leveraging multiple forward passes, this methodology significantly reduces memory and computational requirements. This makes it particularly well-suited for continuous health monitoring applications, where efficiency is paramount. The validation of this approach using multiple datasets not only underscores its effectiveness but also demonstrates its practical applicability in real-world scenarios, highlighting significant improvements in performance and efficiency.
The narrative then shifts towards Federated Learning (FL), focusing on its application in ensuring data privacy while maintaining the efficiency of seizure detection models. A decentralized FL framework is presented, specifically designed to tackle the challenges associated with non-independent and identically distributed (Non-IID) data across medical facilities. The framework incorporates adaptive ensemble learning and a strategic deployment phase, facilitating the creation of personalized, efficient, and privacy-compliant seizure detection models suitable for wearable technology.
Together, these advances contribute significantly to the field of biomedical engineering, particularly within the scope of wearable health monitoring systems for epilepsy. By addressing essential factors such as accuracy, privacy, and user accessibility, the presented methodologies indicate the arrival of next-generation wearable devices, aimed at enhancing the quality of life for individuals prone to epileptic seizures.ES
Hyperdimensional computing for biosignal monitoring: Applications for epilepsy detection
No full textHyperdimensional (HD) computing is a novel approach to machine learning inspired by neuroscience, which uses vectors in a hyper-dimensional space to represent data and models. This approach has gained significant interest in recent years with applications in various domains, one of which being biomedical applications. However, wearable biomedical applications pose a broad range of challenges for hyperdimensional computing that must be tackled before its potentially widespread adoption. I focus on epilepsy detection as a use case and use it for improving and testing the proposed HD computing methods, as it is a chronic neurological disorder that affects a significant portion (0.6 to 0.8%) of the human population and imposes severe risks in the daily life of patients. Despite advances in machine learning and Internet of Things (IoT), small and non-stigmatizing wearable devices for continuous monitoring and detection in outpatient environments are not yet widely available. Thus, the main motivation of my thesis was twofold; first, to explore the advantages and limitations of hyperdimensional computing for biosignal monitoring, and second, to develop new approaches for epilepsy detection.
I demonstrate and develop additional aspects in which HD computing, and the way its models are built and stored, can be used to understand further, compare, and create more complex machine-learning models for epilepsy detection. These possibilities are not feasible with other state-of-the-art models, such as random forests or neural networks. More specifically, I propose new ways to improve two main parts of the HD computing workflow: encoding and learning. Different methods to encode three-dimensional sets of information (features with spatial and temporal information) are proposed and discussed. Next, due to the highly personalized nature of epileptic seizures and their unbalanced nature, learning is improved by proposing a new multi-centroid learning approach.
Then I study the interplay between personalized and generalized models. The process of creation of generalized models from personalized ones is studied, which is interesting for future distributed learning applications. Next, I show that HD computing enables combining personalized and generalized models forming hybrid models, resulting in increasingly performant epilepsy detection. Finally, such models are used to test the knowledge transfer between different datasets, making a first step towards the integration of knowledge from available epilepsy datasets.
The need for interpretable models and predictions in healthcare applications is paramount, and this PhD thesis demonstrates the possibility of HD computing for visualizing prediction decisions in time, per features, and also per channels. Also, the process of feature and channel selection using HD computing encoding is explored.
In the end, I led the development of HDTorch, an open-source PyTorch-based library that enables much faster exploration and development of HD computing algorithms.
Overall, I demonstrate how HD computing can help bring wearable and interpretable healthcare systems closer to reality and patients' everyday life. Despite using epilepsy as a representative use case, all the work proposed is easily translatable to other biomedical signals and applications. Thus, I believe it can inspire and foster further improvements in the hyperdimensional computing field and in wearable healthcare applications.ES
Self-Aware Machine Learning for Chronic Pathology Monitoring on Wearable Devices
No full textRemote health monitoring has attracted a lot of attention over the past decades to provide the opportunity for early detection of pathological health conditions. This early detection improves the quality of life for the patients and significantly reduces the burden on their family members. Moreover, this improvement reduces the socioeconomic consequences that are caused due to the disability of patients to work despite their health conditions.
Wearable technologies offer a promising solution in pervasive health monitoring by relaxing the constraints concerning time and location. There are two major challenges in developing these technologies. The first challenge is real-time monitoring of the patients to fulfill early detection of life-threatening conditions using advanced machine learning techniques. As a solution for this issue, in this thesis, I considered a paradigm shift toward self-aware approaches and frameworks, which resulted in less energy consumption and longer battery lifetime, hence enabling ambulatory patient monitoring. I provided the possibility of switching between energy-efficient and high-performance modes in these systems. The second challenge is improving performance while guaranteeing the accuracy of the system. In this work, I have addressed this second challenge by combining multi-parametric bio-signals analysis and personalizing the detection for each patient. Throughout this thesis, I focused on epileptic seizure detection systems as a principal case study to analyze and evaluate the proposed techniques.ES
Marco de emulación térmica para MPSoCs basado en NoC con DVFS
Full text linkMáster en Investigación en Informática, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, curso 2008-2009El escalado en la tecnología impone un creciente stress sobre la temperatura en el diseño de circuitos digitales debido a la densidad de transistores, especialmente en sistemas de alta integración como son los Multi-Processor System-on-Chip (MPSoCs).
De esta manera, los diseños que tengan en cuenta la temperatura son necesarios y deben ser implementados en las fases tempranas del diseño de los MPSoCs para evitar iteraciones y retrasos en el desarrollo del producto final para el consumidor. En este proyecto presentamos una novel infraestructura de hardware que provee control térmico de las arquitecturas MPSoC, la cual está basada en explotar la interconexión de la NoC del sistema base como un componente activo que comunique y coordine entre los sensores de temperatuea colocados dentro del chip, de esta manera se puede monitorizar globalmente la temperatura actual del sistema. Entonces, una thermal management unit y unos clock frequency controllers son incluidos como parte de la infraestructura térmica activa basada en NoC para ajustar la frecuencia y el voltaje de los elementos de proceso de acuerdo con los requerimientos de temperatura de cada diseño de los MPSoC en runtime. Mostramos resultados experimentales de la aplicación de la propuesta infraestructura de manejo térmico basada en NoC activa para implentar unas efectivas políticas de control térmico de ámbito global para un MPSoC de 4 cores de la vida real, corriendo benchmarks de procesado de video de la vida real, emulados en un marco de emulación térmica basado en una FPGA. Además, debido al mejor balance térmico de nuestro propuesto control térmico basado en NoC activa, el rendimiento del MPSoC mejora al menos un 40% y consigue un 45% de ahorro de energía respecto a las aproximaciones de control térmico con DVFS local.
[ABSTRACT]
Technology scaling imposes an ever increasing temperature stress on digital circuit design due to transistor density, especially on highly integrated systems, such as Multi-Processor Systems-on-Chip (MPSoCs).
Therefore, temperatureaware design is mandatory and should be performed at the early design stages of MPSoCs to avoid iterations and delays in the deployment of final consumer products. In this proyect we present a novel hardware infrastructure to provide thermal control of MPSoC architectures, which is based on exploiting the NoC interconnects of the baseline system as an active component to communicate and coordinate between temperature sensors scattered around the chip, in order to globally monitor the actual temperature of the system. Then, a thermal management unit and clock frequency controllers are included as part of the active NoC-based thermal control infrastructure to adjust the frequency and voltage of the processing elements according to the temperature requirements of each MPSoC design at runtime. We show experimental results of the application of the proposed active NoC-based thermal management infrastructure to implement effective global temperature control policies for a real-life 4-core MPSoC, running real-life video processing benchmarks, emulated on an FPGA-based thermal emulation framework. Furthermore, due to the better thermal balancing of our proposed active NoC-based thermal control, the MPSoC performance improves almost 40% and achieves 45% energy savings with respect to local DVFS thermal control approaches.Depto. de Arquitectura de Computadores y AutomáticaFac. de InformáticaTRUEunpu
- …
