arXiv subjects

Sangyeop Lee

Publications and source records attributed to Sangyeop Lee.

15 recordsLinked to original sources

arXiv2026

Solving the Peierls-Boltzmann transport equation with matrix product states

The Peierls-Boltzmann transport equation (PBE), which governs non-equilibrium phonon transport, suffers from the curse of dimensionality due to its high-dimensional phase space including both real and modal spaces. We explore the use of matrix product states (MPS) for numerical simulation of the PBE. We show that an MPS configuration based on scattering events combined with a dimensionless form of the solution can drastically increase the locality of correlations between tensors in the MPS representation, enhancing its effectiveness in dimension reduction. We further examine the effects of index ordering in an MPS and find that the highest locality is achieved when tensor chains associated with both real and modal spaces are connected from the coarsest grid to each other in the center of the MPS. Using this optimal configuration and a solver inspired by the density matrix renormalization group, we solve the PBE discretized by a finite volume method (FVM). The solution is obtained for crystalline silicon across ballistic, quasi-ballistic, and diffusive transport regimes. An MPS truncated to the compression ratio of $10^{-3}$ suffices to reproduce reference solutions with high fidelity. The computational cost scales sublinearly with the number of grid points in both real and modal spaces, achieving roughly an order of magnitude reduction in computational time compared to the FVM with sparse matrix operation.

cond-mat.mes-hall↗

arXiv2025

Grain boundary-limited thermal transport in suspended thin graphite across an unexplored thickness regime

We present systematic thermal conductivity measurements of suspended thin graphite ribbons, 234-527 nm thick, using a four-probe 3-omega method. Unlike recent reports of phonon hydrodynamics and exceptionally high thermal conductivity in micrometer-thick graphite (Science, 2020),we observe significantly lower thermal conductivity and no signatures of collective phonon flow in this intermediate thickness regime. Instead, our measured thermal conductivity lies between few-layer graphene and bulk graphite.These results agree with a first-principles-informed Peierls-Boltzmann transport model with spatially resolved Monte Carlo sampling. Additionally, the temperature for the peak thermal conductivity shifts lower with increasing thickness, due to the interplay of phonon-boundary and phonon-isotope scattering. Incorporating grain boundary scattering into simulations is necessary to replicate the experimental trends. These findings delineate the boundary between ballistic, hydrodynamic, and diffusive transport regimes in graphite, and underscore the dominant role of disorder and geometry in phonon transport in quasi-two-dimensional materials, offering insights for nanoscale thermal management.

cond-mat.mes-hall↗

arXiv2025

Déjà Vu: Efficient Video-Language Query Engine with Learning-based Inter-Frame Computation Reuse

Recently, Video-Language Models (VideoLMs) have demonstrated remarkable capabilities, offering significant potential for flexible and powerful video query systems. These models typically rely on Vision Transformers (ViTs), which process video frames individually to extract visual embeddings. However, generating embeddings for large-scale videos requires ViT inferencing across numerous frames, posing a major hurdle to real-world deployment and necessitating solutions for integration into scalable video data management systems. This paper introduces Déjà Vu, a video-language query engine that accelerates ViT-based VideoLMs by reusing computations across consecutive frames. At its core is ReuseViT, a modified ViT model specifically designed for VideoLM tasks, which learns to detect inter-frame reuse opportunities, striking an effective balance between accuracy and reuse. Although ReuseViT significantly reduces computation, these savings do not directly translate into performance gains on GPUs. To overcome this, Déjà Vu integrates memory-compute joint compaction techniques that convert the FLOP savings into tangible performance gains. Evaluations on three VideoLM tasks show that Déjà Vu accelerates embedding generation by up to a 2.64x within a 2% error bound, dramatically enhancing the practicality of VideoLMs for large-scale video analytics.

cs.DC↗

arXiv2024

ONNXim: A Fast, Cycle-level Multi-core NPU Simulator

As DNNs are widely adopted in various application domains while demanding increasingly higher compute and memory requirements, designing efficient and performant NPUs (Neural Processing Units) is becoming more important. However, existing architectural NPU simulators lack support for high-speed simulation, multi-core modeling, multi-tenant scenarios, detailed DRAM/NoC modeling, and/or different deep learning frameworks. To address these limitations, this work proposes ONNXim, a fast cycle-level simulator for multi-core NPUs in DNN serving systems. It takes DNN models represented in the ONNX graph format generated from various deep learning frameworks for ease of simulation. In addition, based on the observation that typical NPU cores process tensor tiles from on-chip scratchpad memory with deterministic compute latency, we forgo a detailed modeling for the computation while still preserving simulation accuracy. ONNXim also preserves dependencies between compute and tile DMAs. Meanwhile, the DRAM and NoC are modeled in cycle-level to properly model contention among multiple cores that can execute different DNN models for multi-tenancy. Consequently, ONNXim is significantly faster than existing simulators (e.g., by up to 384x over Accel-sim) and enables various case studies, such as multi-tenant NPUs, that were previously impractical due to slow speed and/or lack of functionalities. ONNXim is publicly available at https://github.com/PSAL-POSTECH/ONNXim.

cs.AR↗

arXiv2024

NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing

Modern transformer-based Large Language Models (LLMs) are constructed with a series of decoder blocks. Each block comprises three key components: (1) QKV generation, (2) multi-head attention, and (3) feed-forward networks. In batched processing, QKV generation and feed-forward networks involve compute-intensive matrix-matrix multiplications (GEMM), while multi-head attention requires bandwidth-heavy matrix-vector multiplications (GEMV). Machine learning accelerators like TPUs or NPUs are proficient in handling GEMM but are less efficient for GEMV computations. Conversely, Processing-in-Memory (PIM) technology is tailored for efficient GEMV computation, while it lacks the computational power to handle GEMM effectively. Inspired by this insight, we propose NeuPIMs, a heterogeneous acceleration system that jointly exploits a conventional GEMM-focused NPU and GEMV-optimized PIM devices. The main challenge in efficiently integrating NPU and PIM lies in enabling concurrent operations on both platforms, each addressing a specific kernel type. First, existing PIMs typically operate in a "blocked" mode, allowing only either NPU or PIM to be active at any given time. Second, the inherent dependencies between GEMM and GEMV in LLMs restrict their parallel processing. To tackle these challenges, NeuPIMs is equipped with dual row buffers in each bank, facilitating the simultaneous management of memory read/write operations and PIM commands. Further, NeuPIMs employs a runtime sub-batch interleaving technique to maximize concurrent execution, leveraging batch parallelism to allow two independent sub-batches to be pipelined within a single NeuPIMs device. Our evaluation demonstrates that compared to GPU-only, NPU-only, and a naïve NPU+PIM integrated acceleration approaches, NeuPIMs achieves 3$\times$, 2.4$\times$ and 1.6$\times$ throughput improvement, respectively.

cs.AR↗

arXiv2023

Non-equilibrium Thermal Resistance of Interfaces Between III-V Compounds

Interfacial thermal resistance has been often estimated and understood using the Landauer formalism that assumes incident phonons with equilibrium distribution. However, previous studies suggest that phonons are out-of-equilibrium near the interface because of the heat flow through the leads and the scattering of phonons by the interface. In this paper, we report a systematic study on how vibrational spectra mismatch affects the degree of phonon non-equilibrium near an interface, how fast it is relaxed as the phonons diffuse into a lead, and the overall interfacial thermal resistance from the non-equilibrium phonons. Our discussion is based on the solution of the Peierls-Boltzmann transport equation with ab initio inputs for 36 interfaces between semi-infinite group-III (Al, Ga, In) and group-V (P, As, Sb) compound semiconductor leads. The simulation reveals that the non-equilibrium phonons cause significant interfacial thermal resistance for all 36 interfaces, making the overall interfacial thermal resistance two to three times larger than that predicted by the Landauer formalism. We observe a clear trend that the degree of phonon non-equilibrium near an interface and the interfacial thermal resistance from the non-equilibrium phonons increase as the mismatch of the Debye temperature of two lead materials increases. This contrasts with Landauer formalism's predictions, which show no correlation with the Debye temperature mismatch. The relaxation length of the phonon non-equilibrium varies significantly from 50nm to 1.5um depending on the combination of the lead materials. The relaxation length is proportional to the phonon mean free path of the corresponding lead material but also largely depends on the material in the opposite lead. This suggests the relaxation length cannot be considered an intrinsic property of the corresponding lead material.

cond-mat.mes-hall↗

arXiv2023

Thermal resistance from non-equilibrium phonons at Si-Ge interface

As nanostructured devices become prevalent, interfaces often play an important role in thermal transport phenomena. However, interfacial thermal transport remains poorly understood due to complex physics across a wide range of length scales from atomistic to microscale. Past studies on interfacial thermal resistance have focused on interface-phonon scattering at the atomistic scale but overlooked the complex interplay of phonon-interface and phonon-phonon scattering at microscale. Here, we use the Peierls-Boltzmann transport equation to show that the resistance from the phonon-phonon scattering of non-equilibrium phonons near a Si-Ge interface is much larger than that directly caused by the interface scattering. We report that non-equilibrium in phonon distribution leads to significant entropy generation and thermal resistance upon three-phonon scattering by the Boltzmann's H-theorem. The physical origin of non-equilibrium phonons in Ge is explained with the mismatch of phonon dispersion, density-of-states, and group velocity, which serve as general guidance for estimating the non-equilibrium effect on interfacial thermal resistance. Our study bridges a gap between atomistic scale and less studied microscale phenomena, providing comprehensive understanding of overall interfacial thermal transport and the significant role of phonon-phonon scattering.

cond-mat.mes-hall↗

arXiv2021

Ab initio phonon transport across grain boundaries in graphene using machine learning based on small dataset

Establishing the structure-property relationship for grain boundaries (GBs) is critical for developing next generation functional materials, but has been severely hampered due to its extremely large configurational space. Atomistic simulations with low computational cost and high predictive power are strongly desirable, but the conventional simulations using empirical interatomic potentials and density functional theory suffer from the lack of predictive power and high computational cost, respectively. A machine learning interatomic potential (MLIP) recently emerged but often requires an extensive size of the training dataset, making it a less feasible approach. Here we demonstrate that an MLIP trained with a rationally designed small training dataset can predict thermal transport across GBs in graphene with ab initio accuracy at an affordable computational cost. In particular, we employed a rational approach based on the structural unit model to find a small set of GBs that can represent the entire configurational space and thus can serve as a cost-effective training dataset for the MLIP. Only 5 GBs were found to be enough to represent the entire configurational space of graphene GBs. Using the atomistic Green's function approach and the MLIP, we revealed that the structure-thermal resistance relation in graphene does not follow the common understanding that large dislocation density causes larger thermal resistance. In fact, thermal resistance is nearly independent of dislocation density at room temperature and is higher when the dislocation density is small at sub-room temperature. We explain this intriguing behavior with the buckling near a GB causing a strong scattering of flexural phonon modes.

cond-mat.mtrl-sci↗

arXiv2021

Transient Hydrodynamic Lattice Cooling by Picosecond Laser Irradiation of Graphite

Recent theories and experiments have suggested hydrodynamic phonon transport features in graphite at unusually high temperatures. Here, we report a pico-second pump-probe thermal reflectance measurement of heat pulse propagation in graphite. The measurement results reveal transient lattice cooling near the adiabatic center of a 15 $μ$m diameter ring-shape pump beam at temperatures between 80 and 120 K. While such lattice cooling has not been reported in recent diffraction measurements of second sound in graphite, the observation here is consistent with both hydrodynamic phonon transport theory and prior heat pulse measurements of second sound in bulk sodium fluoride.

cond-mat.mtrl-sci↗

arXiv2020

Mie Scattering of Phonons by Point Defects in IV-VI Semiconductors PbTe and GeTe

Point defects in solids such as vacancy and dopants often cause large thermal resistance. Because the lattice site occupied by a point defect has a much smaller size than phonon wavelengths, the scattering of thermal acoustic phonons by point defects in solids has been widely assumed to be the Rayleigh scattering type. In contrast to this conventional perception, using an ab initio Green's function approach, we show that the scattering by point defects in PbTe and GeTe exhibits Mie scattering characterized by a weaker frequency dependence of the scattering rates and highly asymmetric scattering phase functions. These unusual behaviors occur because the strain field induced by a point defect can extend for a long distance much larger than the lattice spacing. Because of the asymmetric scattering phase functions, the widely used relaxation time approximation fails with an error of ~20% at 300K in predicting lattice thermal conductivity when the vacancy fraction is 1%. Our results show that the phonon scattering by point defects in IV-VI semiconductors cannot be described by the simple kinetic theory combined with Rayleigh scattering.

physics.comp-ph↗

arXiv2019

Machine-learning based interatomic potential for phonon transport in perfect crystalline Si and crystalline Si with vacancies

We report that single interatomic potential, developed using Gaussian regression of density functional theory calculation data, has high accuracy and flexibility to describe phonon transport with ab initio accuracy in two different atomistic configurations: perfect crystalline Si and crystalline Si with vacancies. The high accuracy of second- and third-order force constants from the Gaussian approximation potential (GAP) are demonstrated with phonon dispersion, Grüneisen parameter, three-phonon scattering rate, phonon-vacancy scattering rate, and thermal conductivity, all of which are very close to the results from density functional theory calculation. We also show that the widely used empirical potentials (Stillinger-Weber and Tersoff) produce much larger errors compared to the GAP. The computational cost of GAP is higher than the two empirical potentials, but five orders of magnitude lower than the density functional theory calculation. Our work shows that GAP can provide a new opportunity for studying phonon transport in partially disordered crystalline phases with the high predictive power of ab initio calculation but at a feasible computational cost.

cond-mat.mtrl-sci↗

arXiv2019

Hydrodynamic Phonon Transport: Past, Present, and Prospect

The hydrodynamic phonon transport was studied several decades ago for verifying the quantum theory of lattice thermal transport. Recent prediction of significant hydrodynamic phonon transport in graphitic materials shows its practical importance for high thermal conductivity materials and brought a renewed attention. As the study on this topic has been inactive to some extent for several decades, we aim at providing a brief overview of the past studies as well as very recent studies. The topics we discuss in this chapter include the collective motion of phonons, several approaches to solve the Peierls-Boltzmann transport equation for hydrodynamic phonon transport, the role of normal scattering for thermal resistance, and the propagation of second sound. Then, we close this chapter with our perspectives for the future studies and the practical implication of hydrodynamic phonon transport.

cond-mat.mes-hall↗

arXiv2019

Crossover of ballistic, hydrodynamic, and diffusive phonon transport in suspended graphene

Hydrodynamic phonon transport was recently predicted as an important regime for phonon transport in graphitic materials. Many of past studies on hydrodynamic phonon transport have focused on the cases where the hydrodynamic regime predominates over other regimes such that hydrodynamic features can be clearly observed. However, this often requires stringent conditions of temperature and sample size. In many cases, a single regime does not dominate other regimes, but all three regimes - ballistic, hydrodynamic, and diffusive regimes - exist to some extent. Here we assess the extent of three regimes by comparing momentum destruction rates by three different mechanisms, each of which represents a different regime: diffuse boundary scattering without internal phonon scattering (ballistic regime), diffuse boundary scattering combined with normal scattering (hydrodynamic regime), and umklapp scattering (diffusive regime). We solve the Peierls-Boltzmann equation with an ab initio full scattering matrix using a deviational Monte Carlo method. We sample distribution functions of ballistic and scattered particles separately, and thereby compare the momentum destruction rates by the three different mechanisms. Using this framework, we discuss a well-known phenomenon of ballistic-to-hydrodynamic crossover, called phonon Knudsen minimum.

cond-mat.mes-hall↗

arXiv2018

Thermal resistance by transition between collective and non-collective phonon flows in graphitic materials

Phonons in graphitic materials exhibit strong normal scattering (N-scattering) compared to umklapp scattering (U-scattering). The strong N-scattering cause collective phonon flow, unlike the relatively common cases where U-scattering is dominant. If graphitic materials have finite size and contact with hot and cold reservoirs emitting phonons with non-collective distribution, N-scattering change the non-collective phonon flow to the collective phonon flow near the interface between graphitic material and a heat reservoir. We study the thermal resistance by N-scattering during the transition between non-collective and collective phonon flows. Our Monte Carlo solution of Peierls-Boltzmann transport equation shows that the N-scattering in graphitic materials reduce heat flux from the ballistic case by around 15%, 30%, and 40% at 100, 200, and 300 K, respectively. This is significantly larger than ~ 5% reduction of Debye crystal with similar Debye temperature (~ 2300 K). We associate the large reduction of heat flux by N-scattering with the non-linear dispersion and multiple phonon branches with different group velocities of graphitic materials.

cond-mat.mtrl-sci↗

arXiv2014

A Comprehensive Review of Heat Transfer in Thermoelectric Materials and Devices

Solid-state thermoelectric devices are currently used in applications ranging from thermocouple sensors to power generators in satellites, to portable air-conditioners and refrigerators. With the ever-rising demand throughout the world for energy consumption and CO2 reduction, thermoelectric energy conversion has been receiving intensified attention as a potential candidate for waste-heat harvesting as well as for power generation from renewable sources. Efficient thermoelectric energy conversion critically depends on the performance of thermoelectric materials and devices. In this review, we discuss heat transfer in thermoelectric materials and devices, especially phonon engineering to reduce the lattice thermal conductivity of thermoelectric materials, which requires a fundamental understanding of nanoscale heat conduction physics.

cond-mat.mtrl-sci↗