Floating point arithmetic remains expensive on FPGA platforms due to wide datapaths and normalization logic, motivating alternative representations that preserve dynamic range at lower cost. This work introduces the Hybrid Residue Floating Numerical Architecture (HRFNA), a unified arithmetic system that combines carry free residue channels with a lightweight floating point scaling factor. We develop the full mathematical framework, derive bounded error normalization rules, and present FPGA optimized microarchitectures for modular multiplication, exponent management, and hybrid reconstruction. HRFNA is implemented on a Xilinx ZCU104, with Vitis simulation, RTL synthesis, and on chip ILA traces confirming cycle accurate correctness. The architecture achieves over 2.1 times throughput improvement and 38-52 percent LUT reduction compared to IEEE 754 single precision baselines while maintaining numerical stability across long iterative sequences. These results demonstrate that HRFNA offers an efficient and scalable alternative to floating point computation on modern FPGA devices.

Quantum circuit simulation remains essential for developing and validating quantum algorithms, especially as current quantum hardware is limited in scale and quality. However, the growing diversity of simulation methods and software tools creates a high barrier to selecting the most suitable backend for a given circuit. We introduce Maestro, a unified interface for quantum circuit simulation that integrates multiple simulation paradigms - state vector, MPS, tensor network, stabilizer, GPU-accelerated, and p-block methods - under a single API. Maestro includes a predictive runtime model that automatically selects the optimal simulator based on circuit structure and available hardware, and applies backend-specific optimizations such as multiprocessing, GPU execution, and improved sampling. Benchmarks across heterogeneous workloads demonstrate that Maestro outperforms individual simulators in both single-circuit and large batched settings, particularly in high-performance computing environments. Maestro provides a scalable, extensible platform for quantum algorithm research, hybrid quantum-classical workflows, and emerging distributed quantum computing architectures.

This paper discusses the challenges encountered when analyzing the energy efficiency of synthetic benchmarks and the Gromacs package on the Fritz and Alex HPC clusters. Experiments were conducted using MPI parallelism on full sockets of Intel Ice Lake and Sapphire Rapids CPUs, as well as Nvidia A40 and A100 GPUs. The metrics and measurements obtained with the Likwid and Nvidia profiling tools are presented, along with the results. The challenges and pitfalls encountered during experimentation and analysis are revealed and discussed. Best practices for future energy efficiency analysis studies are suggested.

Causal Representation Learning (CRL) aims to uncover the data-generating process and identify the underlying causal variables and relations, whose evaluation remains inherently challenging due to the requirement of known ground-truth causal variables and causal structure. Existing evaluations often rely on either simplistic synthetic datasets or downstream performance on real-world tasks, generally suffering a dilemma between realism and evaluative precision. In this paper, we introduce a new benchmark for CRL using high-fidelity simulated visual data that retains both realistic visual complexity and, more importantly, access to ground-truth causal generating processes. The dataset comprises around 200 thousand images and 3 million video frames across 24 sub-scenes in four domains: static image generation, dynamic physical simulations, robotic manipulations, and traffic situation analysis. These scenarios range from static to dynamic settings, simple to complex structures, and single to multi-agent interactions, offering a comprehensive testbed that hopefully bridges the gap between rigorous evaluation and real-world applicability. In addition, we provide flexible access to the underlying causal structures, allowing users to modify or configure them to align with the required assumptions in CRL, such as available domain labels, temporal dependencies, or intervention histories. Leveraging this benchmark, we evaluated representative CRL methods across diverse paradigms and offered empirical insights to assist practitioners and newcomers in choosing or extending appropriate CRL frameworks to properly address specific types of real problems that can benefit from the CRL perspective. Welcome to visit our: Project page:this https URL, Dataset:this https URL.

Signature-based methods have recently gained significant traction in machine learning for sequential data. In particular, signature kernels have emerged as powerful discriminators and training losses for generative models on time-series, notably in quantitative finance. However, existing implementations do not scale to the dataset sizes and sequence lengths encountered in practice. We present pySigLib, a high-performance Python library offering optimised implementations of signatures and signature kernels on CPU and GPU, fully compatible with PyTorch's automatic differentiation. Beyond an efficient software stack for large-scale signature-based computation, we introduce a novel differentiation scheme for signature kernels that delivers accurate gradients at a fraction of the runtime of existing libraries.

This work introduces ParamRF: a Python library for efficient, parametric modelling of radio frequency (RF) circuits. Built on top of the next-generation computational library JAX, as well as the object-oriented wrapper Equinox, the framework provides an easy-to-use, declarative modelling interface, without sacrificing performance. By representing circuits as JAX PyTrees and leveraging just-in-time compilation, models are compiled as pure functions into an optimized, algebraic graph. Since the resultant functions are JAX-native, this allows computation on CPUs, GPUs, or TPUs, providing integration with a wide range of solvers. Further, thanks to JAX's automatic differentiation, gradients with respect to both frequency and circuit parameters can be calculated for any circuit model outputs. This allows for more efficient optimization, as well as exciting new analysis opportunities. We showcase ParamRF's typical use-case of fitting a model to measured data via its built-in fitting engines, which include classical optimizers like L-BFGS and SLSQP, as well as modern Bayesian samplers such as PolyChord and BlackJAX. The result is a flexible framework for frequency-domain circuit modelling, fitting and analysis.

In this paper, a parallel symmetric eigensolver with very small matrices in massively parallel processing is considered. We define very small matrices that fit the sizes of caches per node in a supercomputer. We assume that the sizes also fit the exa-scale computing requirements of current production runs of an application. To minimize communication time, we added several communication avoiding and communication reducing algorithms based on Message Passing Interface (MPI) non-blocking implementations. A performance evaluation with up to full nodes of the FX10 system indicates that (1) the MPI non-blocking implementation is 3x as efficient as the baseline implementation, (2) the hybrid MPI execution is 1.9x faster than the pure MPI execution, (3) our proposed solver is 2.3x and 22x faster than a ScaLAPACK routine with optimized blocking size and cyclic-cyclic distribution, respectively.

BlackJAX is a library implementing sampling and variational inference algorithms commonly used in Bayesian computation. It is designed for ease of use, speed, and modularity by taking a functional approach to the algorithms' implementation. BlackJAX is written in Python, using JAX to compile and run NumpPy-like samplers and variational methods on CPUs, GPUs, and TPUs. The library integrates well with probabilistic programming languages by working directly with the (un-normalized) target log density function. BlackJAX is intended as a collection of low-level, composable implementations of basic statistical 'atoms' that can be combined to perform well-defined Bayesian inference, but also provides high-level routines for ease of use. It is designed for users who need cutting-edge methods, researchers who want to create complex sampling methods, and people who want to learn how these work.

Differentiable numerical simulations of physical systems have gained rising attention in the past few years with the development of automatic differentiation tools. This paper presents JAX-SSO, a differentiable finite element analysis solver built with JAX, Google's high-performance computing library, to assist efficient structural design in the built environment. With the adjoint method and automatic differentiation feature, JAX-SSO can efficiently evaluate gradients of physical quantities in an automatic way, enabling accurate sensitivity calculation in structural optimization problems. Written in Python and JAX, JAX-SSO is naturally within the machine learning ecosystem so it can be seamlessly integrated with neural networks to train machine learning models with inclusion of physics. Moreover, JAX-SSO supports GPU acceleration to further boost finite element analysis. Several examples are presented to showcase the capabilities and efficiency of JAX-SSO: i) shape optimization of grid-shells and continuous shells; ii) size (thickness) optimization of continuous shells; iii) simultaneous shape and topology optimization of continuous shells; and iv) training of physics-informed neural networks for structural optimization. We believe that JAX-SSO can facilitate research related to differentiable physics and machine learning to further address problems in structural and architectural design.

When estimating quantities and fields that are difficult to measure directly, such as the fluidity of ice, from point data sources, such as satellite altimetry, it is important to solve a numerical inverse problem that is formulated with Bayesian consistency. Otherwise, the resultant probability density function for the difficult to measure quantity or field will not be appropriately clustered around the truth. In particular, the inverse problem should be formulated by evaluating the numerical solution at the true point locations for direct comparison with the point data source. If the data are first fitted to a gridded or meshed field on the computational grid or mesh, and the inverse problem formulated by comparing the numerical solution to the fitted field, the benefits of additional point data values below the grid density will be lost. We demonstrate, with examples in the fields of groundwater hydrology and glaciology, that a consistent formulation can increase the accuracy of results and aid discourse between modellers and observationalists. To do this, we bring point data into the finite element method ecosystem as discontinuous fields on meshes of disconnected vertices. Point evaluation can then be formulated as a finite element interpolation operation (dual-evaluation). This new abstraction is well-suited to automation, including automatic differentiation. We demonstrate this through implementation in Firedrake, which generates highly optimised code for solving Partial Differential Equations (PDEs) with the finite element method. Our solution integrates with dolfin-adjoint/pyadjoint, allowing PDE-constrained optimisation problems, such as data assimilation, to be solved through forward and adjoint mode automatic differentiation.

Modern research in code generators for dense linear algebra computations has shown the ability to produce optimized code with a performance which compares and often exceeds the one of state-of-the-art implementations by domain experts. However, the underlying infrastructure is often developed in isolation making the interconnection of logically combinable systems complicated if not impossible. In this paper, we propose to leverage MLIR as a unifying compiler infrastructure for the optimization of dense linear algebra operations. We propose a new MLIR dialect for expressing linear algebraic computations including matrix properties to enable high-level algorithmic transformations. The integration of this new dialect in MLIR enables end-to-end compilation of matrix computations via conversion to existing lower-level dialects already provided by the framework.

Scientific Computing relies on executing computer algorithms coded in some programming languages. Given a particular available hardware, algorithms speed is a crucial factor. There are many scientific computing environments used to code such algorithms. Matlab is one of the most tremendously successful and widespread scientific computing environments that is rich of toolboxes, libraries, and data visualization tools. OpenCV is a (C++)-based library written primarily for Computer Vision and its related areas. This paper presents a comparative study using 20 different real datasets to compare the speed of Matlab and OpenCV for some Machine Learning algorithms. Although Matlab is more convenient in developing and data presentation, OpenCV is much faster in execution, where the speed ratio reaches more than 80 in some cases. The best of two worlds can be achieved by exploring using Matlab or similar environments to select the most successful algorithm; then, implementing the selected algorithm using OpenCV or similar environments to gain a speed factor.

Iterative methods on irregular grids have been used widely in all areas of comptational science and engineering for solving partial differential equations with complex geometry. They provide the flexibility to express complex shapes with relatively low computational cost. However, the direction of the evolution of high-performance processors in the last two decades have caused serious degradation of the computational efficiency of iterative methods on irregular grids, because of relatively low memory bandwidth. Data compression can in principle reduce the necessary memory memory bandwidth of iterative methods and thus improve the efficiency. We have implemented several data compression algorithms on the PEZY-SC processor, using the matrix generated for the HPCG benchmark as an example. For the SpMV (Sparse Matrix-Vector multiplication) part of the HPCG benchmark, the best implementation without data compression achieved 11.6Gflops/chip, close to the theoretical limit due to the memory bandwidth. Our implementation with data compression has achieved 32.4Gflops. This is of course rather extreme case, since the grid used in HPCG is geometrically regular and thus its compression efficiency is very high. However, in real applications, it is in many cases possible to make a large part of the grid to have regular geometry, in particular when the resolution is high. Note that we do not need to change the structure of the program, except for the addition of the data compression/decompression subroutines. Thus, we believe the data compression will be very useful way to improve the performance of many applications which rely on the use of irregular grids.

Recently, there has been growing interest in using standard language constructs (e.g. C++'s Parallel Algorithms and Fortran's do concurrent) for accelerated computing as an alternative to directive-based APIs (e.g. OpenMP and OpenACC). These constructs have the potential to be more portable, and some compilers already (or have plans to) support such standards. Here, we look at the current capabilities, portability, and performance of replacing directives with Fortran's do concurrent using a mini-app that currently implements OpenACC for GPU-acceleration and OpenMP for multi-core CPU parallelism. We replace as many directives as possible with do concurrent, testing various configurations and compiler options within three major compilers: GNU's gfortran, NVIDIA's nvfortran, and Intel's ifort. We find that with the right compiler versions and flags, many directives can be replaced without loss of performance or portability, and, in the case of nvfortran, they can all be replaced. We discuss limitations that may apply to more complicated codes and future language additions that may mitigate them. The software and Singularity containers are publicly provided to allow the results to be reproduced.

The Python programming language is becoming increasingly popular for scientific applications due to its simplicity, versatility, and the broad range of its libraries. A drawback of this dynamic language, however, is its low runtime performance which limits its applicability for large simulations and for the analysis of large data sets, as is common in astrophysics and cosmology. While various frameworks have been developed to address this limitation, most focus on covering the complete language set, and either force the user to alter the code or are not able to reach the full speed of an optimised native compiled language. In order to combine the ease of Python and the speed of C++, we developed HOPE, a specialised Python just-in-time (JIT) compiler designed for numerical astrophysical applications. HOPE focuses on a subset of the language and is able to translate Python code into C++ while performing numerical optimisation on mathematical expressions at runtime. To enable the JIT compilation, the user only needs to add a decorator to the function definition. We assess the performance of HOPE by performing a series of benchmarks and compare its execution speed with that of plain Python, C++ and the other existing frameworks. We find that HOPE improves the performance compared to plain Python by a factor of 2 to 120, achieves speeds comparable to that of C++, and often exceeds the speed of the existing solutions. We discuss the differences between HOPE and the other frameworks, as well as future extensions of its capabilities. The fully documented HOPE package is available at http://hope.phys.ethz.ch and is published under the GPLv3 license on PyPI and GitHub.

Computational fluid dynamics (CFD) is an important tool for the simulation of the cardiovascular function and dysfunction. Due to the complexity of the anatomy, the transitional regime of blood flow in the heart, and the strong mutual influence between the flow and the physical processes involved in the heart function, the development of accurate and efficient CFD solvers for cardiovascular flows is still a challenging task. In this paper we present lifex-cfd, an open-source CFD solver for cardiovascular simulations based on the lifex finite element library, written in modern C++ and exploiting distributed memory parallelism. We model blood flow in both physiological and pathological conditions via the incompressible Navier-Stokes equations, accounting for moving cardiac valves, moving domains, and transition-to-turbulence regimes. In this paper, we provide an overview of the underlying mathematical formulation, numerical discretization, implementation details and examples on how to use lifex-cfd. We verify the code through rigorous convergence analyses, and we show its almost ideal parallel speedup. We demonstrate the accuracy and reliability of the numerical methods implemented through a series of idealized and patient-specific vascular and cardiac simulations, in different physiological flow regimes. The lifex-cfd source code is available under the LGPLv3 license, to ensure its accessibility and transparency to the scientific community, and to facilitate collaboration and further developments.

There is a continuing interest in using standard language constructs for accelerated computing in order to avoid (sometimes vendor-specific) external APIs. For Fortran codes, the {\tt do concurrent} (DC) loop has been successfully demonstrated on the NVIDIA platform. However, support for DC on other platforms has taken longer to implement. Recently, Intel has added DC GPU offload support to its compiler, as has HPE for AMD GPUs. In this paper, we explore the current portability of using DC across GPU vendors using the in-production solar surface flux evolution code, HipFT. We discuss implementation and compilation details, including when/where using directive APIs for data movement is needed/desired compared to using a unified memory system. The performance achieved on both data center and consumer platforms is shown.

Randomized numerical linear algebra - RandNLA, for short - concerns the use of randomization as a resource to develop improved algorithms for large-scale linear algebra computations. The origins of contemporary RandNLA lay in theoretical computer science, where it blossomed from a simple idea: randomization provides an avenue for computing approximate solutions to linear algebra problems more efficiently than deterministic algorithms. This idea proved fruitful in the development of scalable algorithms for machine learning and statistical data analysis applications. However, RandNLA's true potential only came into focus upon integration with the fields of numerical analysis and "classical" numerical linear algebra. Through the efforts of many individuals, randomized algorithms have been developed that provide full control over the accuracy of their solutions and that can be every bit as reliable as algorithms that might be found in libraries such as LAPACK. Recent years have even seen the incorporation of certain RandNLA methods into MATLAB, the NAG Library, NVIDIA's cuSOLVER, and SciKit-Learn. For all its success, we believe that RandNLA has yet to realize its full potential. In particular, we believe the scientific community stands to benefit significantly from suitably defined "RandBLAS" and "RandLAPACK" libraries, to serve as standards conceptually analogous to BLAS and LAPACK. This 200-page monograph represents a step toward defining such standards. In it, we cover topics spanning basic sketching, least squares and optimization, low-rank approximation, full matrix decompositions, leverage score sampling, and sketching data with tensor product structures (among others). Much of the provided pseudo-code has been tested via publicly available MATLAB and Python implementations.