A5070514518- Parallel Computing and Optimization Techniques
- Advancements in Semiconductor Devices and Circuit Design
- Semiconductor materials and devices
- Ferroelectric and Negative Capacitance Devices
- Advanced Neural Network Applications
- Low-power high-performance VLSI design
- Advanced Memory and Neural Computing
- Advanced Data Storage Technologies
- Cloud Computing and Resource Management
- Computational Physics and Python Applications
- Matrix Theory and Algorithms
- Graph Theory and Algorithms
- Distributed and Parallel Computing Systems
- Tensor decomposition and applications
- Numerical Methods and Algorithms
- Neuroscience and Neural Engineering
- Software Testing and Debugging Techniques
- Machine Learning in Materials Science
- Stochastic Gradient Optimization Techniques
- Flood Risk Assessment and Management
- Hydrology and Watershed Management Studies
- Adversarial Robustness in Machine Learning
- Advanced Graph Neural Networks
- Quantum Computing Algorithms and Architecture
- Radiation Effects in Electronics
ETH Zurich
2019-2025
The ubiquity of accelerators in high-performance computing has driven programming complexity beyond the skill-set average domain scientist. To maintain performance portability future, it is imperative to decouple architecture-specific paradigms from underlying scientific computations. We present Stateful DataFlow multiGraph (SDFG), a data-centric intermediate representation that enables separating program definition its optimization. By combining fine-grained data dependencies with...
Biological neural networks do not only include long-term memory and weight multiplication capabilities, as commonly assumed in artificial networks, but also more complex functions such short-term memory, plasticity, meta-plasticity - all collocated within each synapse. Here, we demonstrate memristive nano-devices based on SrTiO3 that inherently emulate these synaptic functions. These memristors operate a non-filamentary, low conductance regime, which enables stable energy efficient...
We introduce Deep500: the first customizable benchmarking infrastructure that enables fair comparison of plethora deep learning frameworks, algorithms, libraries, and techniques. The key idea behind Deep500 is its modular design, where factorized into four distinct levels: operators, network processing, training, distributed training. Our evaluation illustrates (enables combining different codes) (uses carefully selected metrics). Moreover, fast (incurs negligible overheads), verifiable...
On 18 September 2020, the Karditsa prefecture of Thessaly region (Greece) experienced a catastrophic flood as consequence IANOS hurricane. This intense phenomenon was characterized by rainfall records ranging from 220 mm up to 530 mm, in time interval 15 h. Extended public infrastructure damaged and thousands houses commercial properties were flooded, while four casualties recorded. The aim this study provide forensic research on reconstruction event vicinity city. First, we performed...
The computational efficiency of a state the art ab initio quantum transport (QT) solver, capable revealing coupled electro-thermal properties atomically-resolved nano-transistors, has been improved by up to two orders magnitude through data centric reorganization application. approach yields coarse-and fine-grained data-movement characteristics that can be used for performance and communication modeling, communication-avoidance, dataflow transformations. resulting code tuned top-6 hybrid...
Graph neural networks (GNNs) have shown promise in learning the ground-state electronic properties of materials, subverting ab initio density functional theory (DFT) calculations when underlying lattices can be represented as small and/or repeatable unit cells (i.e., molecules and periodic crystals). Realistic systems are, however, non-ideal generally characterized by higher structural complexity. As such, they require large (10+ Angstroms) thousands atoms to accurately described. At these...
Python, already one of the most popular languages for scientific computing, has made significant inroads in High Performance Computing (HPC). At center Python's ecosystem is NumPy, an efficient implementation multi-dimensional array (tensor) structure, together with basic arithmetic and linear algebra. Compared to traditional HPC languages, relatively low performance Python NumPy spawned research compilers frameworks that decouple compact representation from underlying implementation....
Python has become the de facto language for scientific computing. Programming in is highly productive, mainly due to its rich science-oriented software ecosystem built around NumPy module. As a result, demand support High Performance Computing (HPC) skyrocketed. However, itself does not necessarily offer high performance. In this work, we present workflow that retains Python's productivity while achieving portable performance across different architectures. The workflow's key features are...
Determining I/O lower bounds is a crucial step in obtaining communication-efficient parallel algorithms, both across the memory hierarchy and between processors. Current approaches either study specific algorithms individually, disallow programmatic motifs such as recomputation, or produce asymptotic that exclude important constants. We propose novel approach for precise on general class of programs, which we call Simple Overlap Access Programs (SOAP). SOAP analysis covers wide variety from...
Matrix factorizations are among the most important building blocks of scientific computing. State-of-the-art libraries, however, not communication-optimal, underutilizing current parallel architectures. We present novel algorithms for Cholesky and LU that utilize an asymptotically communication-optimal 2.5D decomposition. first establish a theoretical framework deriving I/O lower bounds linear algebra kernels, then its insights to derive schedules, both communicating N^3/(P*sqrt(M)) elements...
The ubiquity of accelerators in high-performance computing has driven programming complexity beyond the skill-set average domain scientist. To maintain performance portability future, it is imperative to decouple architecture-specific paradigms from underlying scientific computations. We present Stateful DataFlow multiGraph (SDFG), a data-centric intermediate representation that enables separating program definition its optimization. By combining fine-grained data dependencies with...
Dense linear algebra kernels are fundamental components of many scientific computing applications. In this work we present a novel method deriving parallel I/O lower bounds for broad family programs. Based on the X-Partitioning abstraction, our explicitly captures inter-statement dependencies. Applying analysis to LU factorization, derive COnfLUX, an algorithm with cost N3/([EQUATION]) communicated elements per processor - only 1/3× over established bound. We evaluate COnfLUX various problem...
We propose a novel approach to iterated sparse matrix dense multiplication, fundamental computational kernel in scientific computing and graph neural network training. In cases where sizes exceed the memory of single compute node, data transfer becomes bottleneck. An based on multiplication algorithms leads sub-optimal scalability fails exploit sparsity problem. To address these challenges, we decomposing into small number highly structured matrices called arrow matrices, which are connected...
Biological neural networks do not only include long-term memory and weight multiplication capabilities, as commonly assumed in artificial networks, but also more complex functions such short-term memory, plasticity, meta-plasticity - all collocated within each synapse. Here, we demonstrate memristive nano-devices based on SrTiO3 that inherently emulate these synaptic functions. These memristors operate a non-filamentary, low conductance regime, which enables stable energy efficient...
The continuous scaling of metal-oxide-semiconductor field-effect transistors (MOSFETs) has led to device geometries where charged carriers are increasingly confined ever smaller channel cross sections. This development is associated with reduced screening long-range Coulomb interactions. To accurately predict the behavior such ultra-scaled devices, electron-electron (e-e) interactions must be explicitly incorporated in their quantum transport simulation. In this paper, we present an...
Flood modelling is among the most challenging scientific task because it covers a wide area of complex physical phenomena associated with highly uncertain and non-linear processes where development physically interpretive solutions usually suffers from lack recorded data [...]
Graph attention models (A-GNNs), a type of Neural Networks (GNNs), have been shown to be more powerful than simpler convolutional GNNs (C-GNNs). However, A-GNNs are complex program and difficult scale. To address this, we develop novel mathematical formulation, based on tensors that group all the feature vectors, targeting both training inference A-GNNs. The formulation enables straightforward adoption communication-minimizing routines, it fosters optimizations such as vectorization,...