150813 research outputs found
Sort by
Simulating LLM Runtime Latency
No full textLarge Language Models (LLMs) are expensive to run and can incur high latencies. Each LLM application has its own cost and latency targets. For example, AI voice assistants operate under low latency objectives, while large document batch processing jobs are typically cost-sensitive. However, navigating these trade-offs is not trivial, as LLM latency is highly task– specific and depends on factors such as the offered query load, the hardware configurations, request properties, and various model characteristics. To support the user in configuring their deployment according to their application needs, we introduce vLLMSim, an accurate simulator that estimates the latency of a given workload on different hardware configurations. vLLMSim advances two key avenues toward latency-aligned LLM deployments. First, the simulated latency metrics inform the user’s model and hardware choice, so they can use a configuration that is ideal for their workload. Second, our simulator enables researchers to quickly test latency-improving ideas, bypassing the need for time-consuming implementations before validating their effectiveness. In fact, vLLMSim is already used in two research projects with the goal of reducing latency and cost of LLM inference. In this thesis, we show how vLLMSim’s design allows it to accurately support the use cases above, while providing highly accurate runtime predictions. To support hardware exploration without GPU access, vLLMSim provides precomputed performance profiles that are sufficient to accurately simulate the user’s workload. The simulator code can be found here, and the instrumented vLLM code for creating profiles can be found here.M.Eng
Tailoring Complexity of Model-Based Controllers for Legged Robots
No full textHumanoid robots promise human-like mobility, but must manage complex and often conflicting control objectives. While model-based controllers can address these challenges using online optimization, they have high computational demands. Model predictive control (MPC) provides closed-loop stability with online trajectory optimization, but achieving real-time rates is difficult for high-dimensional systems. To mitigate this limitation, most MPC implementations rely on reduced-order models (ROMs) that simplify planning but fail to capture whole-body constraints like joint limits and self-collisions. Reactive whole-body controllers (WBCs) partially address this limitation by projecting ROM trajectories onto some wholebody constraints, but these are restricted to acceleration-level constraints like friction cones and torque limits. This thesis advances humanoid planning and control through a renewed focus on model fidelity, solution accuracy ans solve times with three key contributions. First, we propose the CBF-WBC, which augments reactive WBCs with position constraints using control barrier functions (CBFs), enabling the MIT Humanoid to avoid selfcollisions with minimal computational overhead. As a result, the robot can reactively deviate from infeasible trajectories from a reduced-order MPC. Despite fast solve times below 100 microseconds, conflicts can arise between the reduced-order MPC and the CBF-WBC. To address this, we enable real-time whole-body MPC using the alternating direction method of multipliers (ADMM) to provide low-accuracy solutions at high feedback rates. The controller is reliably deployed on hardware and enables the MIT Humanoid to walk robustly on rough terrains and plan complex crossed-leg and arm motions that enhance stability when recovering from significant disturbances. While low-accuracy solutions often suffice for real-time control, we found that higher accuracy could still improve closed-loop performance if computational speed allows. Building on this insight, we propose a framework to simultaneously optimize solution accuracy and model complexity to maximize closed-loop performance. Instead of planning with a single model that is too complex or too simple, solve times can be reduced by planning over a sequence of models of reducing complexity. We extract ROMs from whole-body dynamics equations and optimize their horizons, discretization timesteps and solution accuracy using blackbox optimization. The optimizer can sacrifice model complexity for additional ADMM iterations, reducing falls by nine-fold and enabling a 2 m/s walking speed on hardware.Ph.D
Opportunities in Advanced Wireless Integrated Circuits
No full textThe continued evolution of wireless communications, novel compact radars, and power electronics has driven demand for high-performance semiconductor materials capable of operating at higher power density, fast switching speeds, and improved efficiency. Gallium Nitride (GaN) has emerged as a leading candidate due to its superior electrical properties compared to traditional silicon (Si), silicon carbide (SiC), and gallium arsenide (GaAs). GaN’s high power density, thermal stability, and high-frequency operation make it an ideal candidate for applications in 5G/6G infrastructure, satellite communications, defense radar, electric vehicles, and power electronics. However, widespread commercial adoption of GaN faces significant barriers, including high production costs, supply chain constraints, and integration challenges within existing silicon-based fabrication processes.
This thesis explores the opportunities and challenges associated with GaN-based integrated circuits (ICs) in the context of advanced wireless systems by utilizing Dr. Eugene Fitzgerald’s innovation framework – Technology, Markets, and Implementation (TMI). A comparative analysis of monolithic vs. board-level GaN integration is conducted. The research highlights that scaling GaN wafer production to approximately 10,000 wafers per year (200mm sized wafers) is necessary to achieve cost-effective monolithic integration, yet current defense-driven demand is insufficient to drive economies of scale. Instead, commercial applications—such as telecommunications, power electronics, and consumer RF devices—are target audiences that can take advantage of monolithic integration in high volume.
The findings indicate that while defense applications have led non-monolithic GaN adoption (that is, discrete GaN transistor adoption), they cannot sustain large-scale production alone due to small volume. The semiconductor industry must navigate manufacturing bottlenecks, cost reduction strategies, and foundry availability to ensure GaN’s transition from a niche, high-cost technology to a commercially viable solution. By mapping the TMI intersections and addressing economic and technical barriers, this thesis provides strategic insights into how GaN technology can achieve scalable production, unlock new market opportunities, and shape the future of advanced wireless integrated circuits.S.M
On approximability of Satisfiable k -CSPs: I
No full textWe consider the P -CSP problem for 3-ary predicates P on satisfiable instances. We show that under certain conditions on P and a ( 1 , s ) integrality gap instance of the P -CSP problem, it can be translated into a dictatorship vs. quasirandomness test with perfect completeness and soundness s + ϵ , for every constant ϵ > 0 . Compared to Ragahvendra (in: Proceedings of the fortieth annual ACM symposium on theory of computing (STOC), pp 245–254, 2008), we do not lose perfect completeness. This is particularly interesting as this test implies new hardness results on satisfiable constraint satisfaction problems, assuming the Rich 2-to-1 Games Conjecture by Braverman et al. (in: Lee JR (ed) Volume 185 of Leibniz international proceedings in informatics (LIPIcs), 27:1–27:20. Schloss Dagstuhl–Leibniz-Zentrum für Informatik, Dagstuhl, 2021b. https://drops.dagstuhl.de/opus/volltexte/2021/13566 ).Our result can be seen as the first step of a potentially long-term challenging program of characterizing optimal inapproximability of every satisfiable k -ary CSP. At the heart of the reduction is our main analytical lemma for a class of 3-ary predicates, which is a generalization of a lemma by Mossel (Geom Funct Anal 19(6):1713–1756, 2010). The lemma and a further generalization of it that we conjecture may be of independent interest
Many-body expansion based machine learning models for octahedral transition metal complexes
No full textGraph-based machine learning (ML) models for material properties show great potential to accelerate virtual high-throughput screening of large chemical spaces. However, in their simplest forms, graph-based models do not include any 3D information and are unable to distinguish stereoisomers such as those arising from different orderings of ligands around a metal center in coordination complexes. In this work we present a modification to revised autocorrelation descriptors, a molecular graph featurization method, for predicting spin state dependent properties of octahedral transition metal complexes (TMCs). Inspired by analytical semi-empirical models for TMCs, the new modeling strategy is based on the many-body expansion (MBE) and allows one to tune the captured stereoisomer information by changing the truncation order of the MBE. We present the necessary modifications to include this approach in two commonly used ML methods, kernel ridge regression and feed-forward neural networks. On a test set composed of all possible isomers of binary TMCs, the best MBE models achieve mean absolute errors (MAEs) of 2.75 kcal mol−1 on spin-splitting energies and 0.26 eV on frontier orbital energy gaps, a 30%–40% reduction in error compared to models based on our previous approach. We also observe improved generalization to previously unseen ligands where the best-performing models exhibit MAEs of 4.00 kcal mol−1 (i.e. a 0.73 kcal mol−1 reduction) on the spin-splitting energies and 0.53 eV (i.e. a 0.10 eV reduction) on the frontier orbital energy gaps. Because the new approach incorporates insights from electronic structure theory, such as ligand additivity relationships, these models exhibit systematic generalization from homoleptic to heteroleptic complexes, allowing for efficient screening of TMC search spaces
Organization Infrastructure for Tokenized Asset Records
No full textThe Tokenized Asset Record (TAR) represents a way to connect existing technology related to tokenized assets and asset schemas to real-world documents that validate the existence of an object. Exactly who should manage TARs and the properties of the related organization schemes remains an open question. Answering this question is crucial to furthering the existing digital economy. While existing solutions have sought to expand digital commerce through pioneering digital clearing houses, little work has explored support for other classes of real-world digitized assets with proof of ownership and existence. The research proposed here seeks to answer this question by suggesting possible solutions and developing a framework for uniformly analyzing the proposals. The research proposes and evaluates three models for the management of TARs. The first is a scheme that involves each industry setting up its own TAR database and managing the system independently from other industries. The second proposes hosting all TARs on a single blockchain. The third argues for an off-chain decentralized platform to host all, akin to the Data Spaces proposed by the European Union. The research finds, based on the proposed criteria, that a decentralized off-chain approach best meets the goals of a TAR management framework.M.Eng
Design of an FSAE Cooling System for Electric In-hub Motors
No full textThis thesis explores the design of water cooled electric in-hub motors for use in high power automotive applications such as Formula SAE Electric student racecars. The main components of the drivetrain of the 4WD electric vehicle are the motor and the motor controller. This thesis focuses on designing a reliable cooling system for the motors to ensure that they operate in an optimal temperature range that increases drivetrain efficiency, prevents catastrophic motor damage, and ultimately improves vehicle performance.
During the design process, extensive heat transfer analysis was conducted for liquid cooling, air cooling, and heat pipes. Additionally, CFD was conducted for various water cooling architectures to determine the influence of coolant flow direction and channel dimensions on cooling performance. Results were subsequently analyzed, plotted, and used in the motor cooling architecture selection process.
After determining the motor cooling sleeve architecture and dimensions, detailed CAD was created for both plastic prototype sleeves as well as metal 3D printed sleeves. Testing demonstrated that the cooling sleeve successfully met its target pressure drop of 3-4PSI. More thorough testing and data collection of the cooling sleeve’s thermal performance is still in progress. While details on the CAD, prototyping, testing & validation, and manufacturing, are outside the scope of this thesis, images are included for reference.S.B
Towards achieving power autonomy in soft-actuated micro aerial robots
No full textMicro aerial robots with insect-like flight capabilities hold immense promise for various applications, including environmental monitoring, precision agriculture, and infrastructure inspection in confined spaces. However, realizing power autonomy in these miniature robotic platforms presents significant challenges due to weight constraints, power density limitations, and inefficient actuation at small scales. This dissertation presents three essential improvements towards achieving power autonomy in soft-actuated micro aerial robots. Our robotic platform is driven by a dielectric elastomer actuator (DEA) and generates lift force through flapping wings, a similar mechanism found in flying insects. First, we implemented a dynamic model to optimize the robot components for pairing with an improved DEA to generate a higher lift force. The robot achieved a peak lift-to-weight ratio of 4.3 and demonstrated a 20-second hovering flight with position and attitude errors smaller than 2.5 cm and 2◦ . Second, we fabricated a lightweight high-voltage boost converter that transformed a 7 V DC input into an AC waveform of 600 V and 400 Hz to drive the actuator. This is the first onboard boost converter that can drive the soft-actuated micro aerial robot to take off, and it represents a substantial achievement in miniaturizing power electronics for microrobots. Third, we took inspiration from the natural autorotation of maple seeds in their slow descent. We implemented the first samara-inspired mechanism on micro aerial robots, enhancing lift generation while maintaining in-flight attitude stability without feedback control. The 1.22-gram vehicle can stably take off in 1 second with a total input thrust of 1 gram-force. These accomplishments provide a pathway towards achieving power autonomy and open opportunities for developing agile, robust, and autonomous micro aerial robots for diverse applications.Ph.D
Systematic Investigation of Silicon Substitution on Single Macromolecule Mechanics
No full textFour unsaturated poly(carbooligosilane)s (P1–P4) were prepared via acyclic diene metathesis polycondensation of new oligosilane diene monomers (1–4). These novel polymers with varying main-chain Si incorporation have high trans internal olefin stereochemistry (ca. 80%) and molecular weights (9500–21,700 g mol–1). Postpolymerization epoxidation converted all alkene moieties to epoxides and rendered the polymers (P5–P8) more electrophilic, which allowed for single-molecule force spectroscopy studies via a modified atomic force microscope setup with a silicon tip and cantilever. The single-chain elasticity of the polycarbooligosilanes decreased with increasing numbers of Si–Si bonds, a finding reproduced by quantum chemical calculations
Analog On-chip Training and Inference with Non-volatile Memory Devices
No full textAs the demand for computation in neural networks continues to rise, conventional computing resources are increasingly constrained by their limited energy efficiency. One promising solution to this challenge is analog in-memory computing (AIMC), which enables efficient matrix-vector multiplications by encoding synaptic weights into the conductance of nonvolatile memory devices. These devices are structured into crossbar arrays. To explore the potential of non-volatile memory devices in AIMC, investigations involve simulating crossbar array operations using IBM’s AIHWKIT. With this tool, I investigate the implementation of various analog computing algorithms, including TikiTaka. AIMC is evaluated for simple MNIST classification tasks and more complex deep learning models, Long Short-Term Memory (LSTM) networks. I demonstrate that devices can be categorized based on their asymmetry and non-linear weight modulation behavior. Performance improvements through the Tikitaka algorithm are observed only when the device provides a sufficient converge-dragging force; otherwise, the algorithm may even degrade performance. I also investigate how pulse-to-pulse noise and device-to-device variability affect system performance, as well as how different peripheral circuit configurations influence the overall behavior. Finally, I propose an Analog Low-Rank Adapter (Analog LoRA) by applying analog computing to the fine-tuning of large language models. I explore the necessary conditions for Analog LoRA to achieve performance comparable to its digital counterpart. Based on these findings, I present design guidelines for effectively applying analog computing to various machine learning tasks on edge devices.S.M