134 research outputs found
Bio-inspired Energy Optimizations to Synchronous Spiking Neural Network Architecture for Reinforcement Learning at Edge Application
No full textReinforcement learning (RL) offers a reward-based model through trial and error within natural contexts, presenting an attractive approach to machine learning due to its simplicity, effectiveness, and similarity to human and animal learning processes. Particularly promising for neuromorphic implementation in autonomous agents' control, RL faces challenges in digital AI accelerators in micro-to-milli-Watt edge applications, where complex features inherited from traditional computing, such as floating-point arithmetic, increase cost and power consumption. Such approaches often struggle to efficiently accommodate biologically inspired models, where learning capacity and accuracy are customized for the task instead of relying on costly upfront hardware resources.
Our research within Energy-Aware Artificial Learning Group targets bio-inspired digital RL hardware model employing simplified integer arithmetic to enhance cost and energy efficiency. The model comprises a 16-node spiking neural network (SNN) template from literature where integer synapse weights adjust through reward and non-reward actions. First architectural optimizations delivered by our group highlighted potential significant energy savings compared to the literature by adopting small integer resolution and simplifying the implementation of Long Term Potentiation (LTP) and Long Term Depression (LTD) operations in Spike-Timing-Dependent-Plasticity (STDP) to track causality over only a few consecutive clock ticks. The implementation focused on a core RL network and required an external processor support for real-time autonomous execution.
This work introduces further bio-inspired optimization to the original RL architecture, providing substantial energy consumption benefits, while at the same time meeting the requirement for real-time autonomous processing with higher accuracy in context-dependent tasks. Synthesis, simulation, and functional validation of real-world implementation on dual-supply Intel MAX10 Field-Programmable Gate Array (FPGA) reveal about an order of magnitude reduction in average power dissipation and even larger benefits in energy consumption compared to state-of-the-art solutions, illustrating the potential of this SNN edge architectural approach.
Index Terms- Reinforcement learning (RL), spiking neural network (SNN), neuromorphic hardware, digital system design, low-power, low-cost, low-energy, leaky integrate and fire (LIF) model, context-dependent task, brain-inspired computing, FPGAs
Smart Power Management for Energy-Autonomous Wearable Health Monitors
No full textThe rise of wearable health monitors (WHMs) has opened new frontiers in personalized healthcare, enabling continuous monitoring of vital signs outside clinical settings. However, the energy demands of these devices present a significant challenge, particularly in achieving long-term, autonomous operation without frequent battery recharge. This thesis explores the potential of Thermoelectric Generators (TEGs) combined with smart power management strategies, particularly Reinforcement Learning (RL), to address this challenge.
The study involved the design and simulation of a TEG-based energy harvesting system integrated with a boost converter, supercapacitor storage, and resistive loads using the MATLAB Simulink toolbox. Two TEG modules were utilized in the design, and they were connected in such a way as to switch between a parallel and a series configuration. The system's performance was evaluated under various configurations and ambient conditions, with a focus on maintaining energy-neutral operation (ENO). Simulation results indicated that the parallel configuration of TEG modules provided better performance at lower ambient temperatures, while the series configuration was more efficient at higher temperatures.
The RL problem was set as a Markovian Decision Process (MDP) consisting of states, action and reward space. A neural framework consisting of six sensory inputs, eight hidden layer and four possible actions was designed for the RL solution. The sensory inputs represented harvested energy from the modelled TEG and the available charge in the supercapacitor storage. A hypothetical analysis of the RL agent's potential behaviour suggested that it would dynamically adjust the system's load states and configuration to optimize energy efficiency based on real-time sensory inputs. Although the RL algorithm has not yet been fully implemented, these findings lay the groundwork for future work aimed at creating a fully autonomous and intelligent power management system for wearable health monitors.
This study contributes to the development of sustainable wearable technologies, offering a promising solution for energy-autonomous WHMs capable of achieving energy neutral operation in diverse environmental conditions
ABACUS: A novel array multiplier-accumulator architecture for low energy applications
No full textDesign and verification of a novel array multiplieraccumulator architecture, named ABACUS, is introduced in this paper. The design priority in this architecture is low energy operation instead of the traditional 'performance-first' approach. ABACUS uses a threshold function to implement multiple fast carry operations in parallel through a cellular array, and therefore significantly deviates from the conventional approaches based on half/full adder or counter building blocks. The ABACUS architecture was formally verified for correct functionality of any unsigned m x n multiplication. Hardware implementation has been validated on FPGAs. The energy-delay advantages were architecturally analyzed over the traditional carry-save array, and circuit implementation focus areas have been accordingly developed. ©2010 IEEE
Recommended from our members
Sustainable power management of microelectronics
No full textThe integration of renewable energy sources and the associated power electronics to notebook power and thermal management systems was investigated in this work. Both thermoelectric (TE) and photovoltaic (PV) energy extraction techniques were explored with the goal of achieving best sustainability. Semi-realistic usage assumptions were taken into account to evaluate the merits of the new schemes, along with relevant metrics that encompassed both performance and energy expenditure.
A Thermoelectric System Solver (“TESS”) was developed, and used to simulate the application of TEs to different platform types. The benefit from the three analyzed TE schemes had strong sensitivity to the platform size and material technology. The Hybrid Thermoelectric Conversion (HTC) scheme offered the best tradeoffs, and demonstrated potential for 10% improvement in system efficiency with common TE components, as measured by Performance/Watt. TE module integration into the heat pipe in a standard Thin and Light mobile system was analyzed through a Finite Element Analysis (FEA) model, which identified acceptable module sizes and placements. Next, Photovoltaic Power Generation (PPG) was added by evaluating an integrated PV sheet at the backside of the notebook lid. Available photovoltaic (PV) power range was quantified per system type under realistic use scenarios. Up to 5% battery life extension was achieved using commonly available PV modules.
Power electronics efficiency was highlighted as one of the fundamental challenges in the integration of renewables into a mobile computing platform. The mobile power architecture was enhanced to store the extracted energy from multiple sources in the existing system battery. Using power conversion efficiency metrics suitable for the renewables application, several fundamental topologies were evaluated for stepping up low voltage TE outputs. A cascaded charge pump scheme with an asynchronous oscillator based control scaled best to the requirements of the application, and yielded positive net power generation to the system battery. The circuit contained no magnetic components to be suitable for large scale integration (LSI). Finally, fundamental conclusions from simulations were verified through hardware prototypes using discrete components
Patient-Centered Design Method for Self-Powered and Cost-Optimized Health Monitors
No full textThe emergence of Wireless Body Area Networks (WBANs) with health monitoring capabilities has revolutionized health care. Implementing fully independent WBAN nodes is important to the long-term viability of this initiative. Regularly recharged and depletable batteries remain a significant impediment in such systems. Energy harvesting (EH) from environmentally clean sources has thus been receiving increasing attention. Nevertheless, the autonomy and optimization of existing WBAN sensor nodes have remained questionable because methods that integrate realistic usage conditions into the design process have been lacking. A plausible method is proposed to establish a framework for designing a sustainable health monitoring node in this work. A Health Monitoring Energy System (HeMeS) tool prototype is consequently developed using comprehensive analytical models and utilized to demonstrate system design space exploration for various patient types, incorporating environmental factors, electronic load activity levels, and system cost/size constraints. It is concluded that the patient-centered system design approach incorporating interactions across transducers, electronics, sensors, user environment and data duty-cycling profiles, is viable, and is in fact appealing in safeguarding truly autonomous and cost-optimal WBANs that are compatible with climate-neutral society
Green PG: A low cost, modular, pedal-powered 5–20 V parallel DC source for mobile computing applications
No full textGreen PG is a pedal powered notebook and cell phone charger, which converts mechanical energy produced by exercise devices into electrical energy with 5 V and 19.5 V DC outlets. Thus, portion of the wasted heat from pedaling is recovered for use. The efficient Green PG architecture is modular, and can be reconfigured for a variety of cost optimized end use scenarios, simultaneously addressing the sustainable energy needs of remote under-developed geographies, long range cyclists, and modern gymnasiums. Performance tests indicate many mobile electronic appliances relevant to daily life can be operated by this green power source. ©2010 IEEE
Design Space Exploration of a Fully Autonomous Health Monitoring WBAN Node with Hybrid Energy Harvesting
No full text© 2021 IEEE.The emergence of WBANs with health monitoring capabilities has greatly revolutionized health care. Enabling WBAN nodes to be fully autonomous is critical in making this thrust sustainable. However, regular charging of batteries in such systems remains a significant inhibitive factor. In this work, Health Monitoring Energy System (HeMeS) tool previously developed by our group using comprehensive analytical models is utilized to study various energy flow scenarios in a particular health monitoring WBAN node powered by a hybrid thermal-vibrational energy harvester. The use of HeMeS for design space exploration is thus demonstrated for various patient categories, incorporating environmental factors, electronic load activity levels, and system cost/size constraints. The described comprehensive system design approach of incorporating transducer, electronics, user environment and data duty-cycling profiles, is demonstrated to be viable and appealing for delivering sustainable WBANs that directly contribute to climate-neutral society without significantly increasing cost
Energy-delay performance of capacitive threshold logic (CTL) circuits for threshold detection
No full textPETAM: Power Estimation Tool for Array Multipliers
No full textIncreasing demand for the mobile, low energy systems has laid emphasis on the development of low power processors. Low power design has to be incorporated into fundamental computation units, such as multipliers. The optimization of the energy-delay product in such low power multipliers will enable energy efficient computation. This study proposes a power estimation tool to analyze different array multiplier architectures, which are most commonly used in such applications. Gate level library design parameters are utilized to derive energy-delay performance for any given set of input vector patterns, and multiplier size. Vector and size dependent factors are therefore clearly identified. Examples are provided from carry save array multiplier (CSAM) and ripple carry array multiplier (RCAM) to demonstrate the capabilities for the tool
- …
