GPU Day 2025

22 May 2025, 08:20 → 23 May 2025, 23:50 Europe/Budapest

HUN-REN Centre

Balázs Kacskovics (Wigner Research Centre for Physics), Gabor Biro (MTA Wigner FK), Gergely Barnafoldi (Wigner RCP RMI of the Hungarian Academy of Sciences)

 

15th GPU Day - Massive parallel computing for science and industrial application

 

The 15th GPU Day will organized by the Wigner Scientific Computation Laboratory, HUN-REN Wigner RCP from May 22-23, 2025 at the HUN-REN Centre (1054 Budapest, Alkotmány utca 29).

The GPU Day is an annually organized international conference series dedicated to massively parallel technologies in scientific and industrial applications for a decade. Its goal is to bring together from academia, developers from industry, and interested students to exchange experiences and learn about future massively parallel technologies. The event provides a unique opportunity to exchange knowledge and expertise on GPU (Graphical Processing Unit), FPGA (Field-Programmable Gate Array), and quantum simulations. As in previous years, we are expecting 80+ participants in person.

 

For the earlier events see: 2024, 2023, 2022, 2021, 2020, 2019, 2018, 2017, 2016, 2015, 2014

 

Participation is free for students, members of academic institutions, research centers, and universities.
The registration fee is 300EUR or 120.000HUF. Participants can pay via bank transfer or card at the registration desk.

 

The conference will be held in offline form and there are limited places to participate personally on-site. We encourage our former, current and future partners to contribute on the conference. Contributions to the conference are welcome.
 

 

Keynote speakers

• Roland Jakab (HUN-REN)
• Sofia Vallecorsa (CERN)
• Michael Doser (CERN)
• István Csabai (ELTE)

 

Important deadlines 

Talk submission deadline: 2025 May 15

 

Sponsors

 

 

More information is available on gpuday.com

---

Organizers:
Gergely Gábor Barnaföldi
Gábor Bíró
Balázs Kacskovics

GPU Day 2025.pdf

qrkod.png

Thursday, 22 May

09:00 → 10:00 Opening: Opening Talk and Welcome by the Director

Conveners: Gergely Barnafoldi (Wigner RCP RMI of the Hungarian Academy of Sciences), Peter Levai (WIGNER RCP)

09:00 Opening

Speakers: Gergely Barnafoldi (Wigner RCP RMI of the Hungarian Academy of Sciences), Peter Levai (WIGNER RCP)

GGBarnafoldi_GPUDay_WSCLAB_2025.pdf

09:15 HUN-REN's AI for Impact Program

Speaker: Roland Jakab (HUN-REN)

GPUday_Jakab Roland_HUN-REN_2025.pptx

09:30 Physics of AI - AI for Sciences

This year's Nobel Prize in Physics came as a bit of a surprise to many, as the awarded topic does not seem to relate to the traditional fields of physics for the untrained observer. In this talk, I will try to explain how models rooted in statistical physics have laid the groundwork for the rapidly developing algorithms of artificial intelligence today and give a brief overview of the current state-of-the-art of AI application in sciences and medicine.

Speaker: István Csabai (ELTE Eötvös University, Dept. of Physics of Complex Systems)

csabai_GPU_Day20250522.pptx

10:00 → 10:45 Session I

10:00 Quantum Machine Learning in HEP. Examples from CERN

Theoretical and algorithmic advances, availability of data, and computing power have opened the door to exceptional perspectives for application of classical Machine Learning in the most diverse fields of science, business and society at large, and notably in High Energy Physics (HEP). In particular, Machine Learning is among the most promising approaches to analyse and understand the data the next generation HEP detectors will produce.

Machine Learning is also an interesting task for quantum devices that can leverage compressed high dimensional representations and use the stochastic nature of quantum measurements as random source. Several architectures are being investigated. Quantum implementations of Boltzmann Machines, classifiers or Auto-Encoders, among the most popular classical approaches, are being proposed for different applications. Born machines are purely quantum models that can generate probability distributions in a unique way, inaccessible to classical computers.
This talk will give an overview of the current state of the art in terms of Machine Learning on quantum computers with focus on their application to HEP.

Speaker: Sofia Vallecorsa (CERN)

CERN QTI Wigner GPU Days May 2025.pdf

10:45 → 11:15 Coffee Break

11:15 → 12:30 Session II

11:15 Quantum Sensing for (low and high energy) particle physics

The seminar will provide a glimpse of some elements of the rapidly evolving field of quantum sensing, specifically focusing on particle physics. Specific approaches involving quantum systems, such as low-dimensional systems or manipulations of ensembles of quantum systems, hold great promise for improving high-energy particle physics detectors, particularly in areas like calorimetry, tracking, and timing. The use of quantum sensors for high-precision measurements, such as precision spectroscopy of novel atomic, molecular or ionic systems, as well as the development of new quantum sensors based on superconducting circuits, ion and particle traps, crystals, and nanomaterials, are equally relevant for low energy particle physics and for fundamental physics.

However, significant advances and improvements in existing or future quantum technologies will be necessary to address such topics related to the dark universe, the detection of relic neutrinos, precision tests of symmetries and of the standard model and probing general foundational issues in physics. The seminar will thus also feature discussions of the Quantum Sensing Initiatives at CERN and the ECFA R&D Roadmap on Quantum Sensing and Advanced Technologies and will discuss options for future collaborations in the context of the recently approved DRD5 implementation of the roadmap.

Speaker: Michael Doser (CERN ? MIT)

GPU_Budapest.pdf

12:00 Batched Line Search Strategy for Navigating through Barren Plateaus in Quantum Circuit Training

Variational quantum algorithms are viewed as promising candidates for demon-
strating quantum advantage on near-term devices. These approaches typically involve
the training of parameterized quantum circuits through a classical optimization loop.
However, they often encounter challenges attributed to the exponentially diminishing
gradient components, known as the barren plateau (BP) problem. This work intro-
duces a novel optimization approach designed to alleviate the adverse effects of BPs
during circuit training. In contrast to conventional gradient descent methods with a
small learning parameter, our approach relies on making a finite hops along the search
direction determined on a randomly chosen subsets of the free parameters. The opti-
mization search direction, together with the range of the search, is determined by the
distant features of the cost-function landscape. This enables the optimization path to
navigate around barren plateaus without the need for external control mechanisms. We
have successfully applied our optimization strategy to quantum circuits comprising 16
qubits and 15000 entangling gates, demonstrating robust resistance against BPs. Addi-
tionally, we have extended our optimization strategy by incorporating an evolutionary
selection framework, enhancing its ability to avoid local minima in the landscape. The
modified algorithm has been successfully utilized in quantum gate synthesis applica-
tions, showcasing a significantly improved efficiency in generating highly compressed
quantum circuits compared to traditional gradient-based optimization approaches

Speaker: Peter Rakyta (Department of Physics of Complex Systems, Eötvös Loránd University)

rakyta_talk.pdf

12:30 → 14:00 Lunch break

14:00 → 15:30 Session III

14:00 Comparison of Classical Digital, Classical Photonic, and Quantum Photonic NNs for Solving PDEs: A Case Study

Physics-Informed Neural Networks (PINNs) aim to solve ordinary differential equations (ODEs) and partial differential equations (PDEs). So far, classical digital NNs have been successfully applied to many PDEs, such as the heat equation, Poisson equation, Navier-Stokes equation, to name a few. Recently, Continuous Variable Quantum Neural Networks (CVQNNs) have also been used to solve PDEs in the context of Quantum Physics-Informed Neural Networks (QPINNs). However, to the best of our knowledge, Classical Photonic networks have not been used for these kinds of tasks (usually various classification tasks are tackled with success). In this study, we plan to compare the method and effectiveness of solving a specific nonlinear PDE, the Burgers equation, in the case of the three networks mentioned above.

Speaker: Dr Peter Kovacs (Wigner RCP)

GPU_day_2025_pkovacs.pdf

14:30 Qubla: a practical language for mixed bit-qubit computations

With the emergence of quantum programming languages and compilers, the necessity for optimization naturally arises. In 2022, we introduced a new quantum language called Qubla, designed to facilitate the generation of quantum counterparts to classical (binary) algorithms. The Qubla compiler constructs a sequence of quantum operators based on a description of mixed classical/quantum computation, "precalculating" all classically computable quantities. However, the resulting chain of operators may rely on an unnecessarily large number of qubits. In this talk, we present methods for simplifying and optimizing the generated quantum chain, and also demonstrate on Shor's algorithm that these are capable of handling practical problems of quantum computing.

Speaker: Gergely Gálfi (ELTE IK)

qubla_gpu2025_v1.pdf

15:00 A line search strategy for training CV variational quantum circuits

Variational quantum algorithms are widely regarded as promising approaches for achieving quantum advantage on near-term hardware. These algorithms typically rely on training parameterized quantum circuits through a classical optimization loop, most often using gradient-based methods with gradients computed via parameter-shift rules. However, in the continuous-variable (CV) model of photonic quantum computing, gradient evaluation poses a significant challenge due to the lack of exact parameter-shift rules. In this work, we introduce an alternative optimization strategy for CV variational quantum circuits based on interpolation techniques.

Speaker: Zoltán Kolarovszki

kz_gpu25_presentation_v2.pdf

15:30 → 16:00 Coffee break

16:00 → 18:30 Session IV

16:00 GPU-cluster-accelerated FVM simulation for the compressible Navier-Stokes equations on unstructured meshes

From the HiDALGO2 Centre of Excellence of the EuroHPC, the RedSim native multi-GPU and vectorized multi-CPU code for simulating the compressible Navier-Stokes equations with unstructured polyhedral meshes has been developed and optimized.
From the HiDALGO2 Centre of Excellence of the EuroHPC, the RedSim native multi-GPU and vectorized multi-CPU code for simulating the compressible Navier-Stokes equations with unstructured polyhedral meshes has been developed and optimized.

In this presentation, we'll go over the different challenges in developing a GPU-oriented CFD Navier-Stokes solver, with manually written CUDA kernels, and the optimization of those kernels using NVIDIA's profiling toolkits. We'll also discuss the development of our MPI + CUDA code, where we leverage the inter-node communication capabilities of MPI in order to run our code on 16, 32, 64, and more GPUs. We'll talk about challenges we've tackled concerning lower cache sizes compared to CPU-s, lower memory environments because of current day VRAM sizes, and intra-node GPU communication via SLI/NVlink.

Finally, we present benchmark results of RedSim when applied to different EuroHPC and the Hungarian HPC Komondor for pipe acoustics problems.

Speakers: Mr Mátyás Constans (Széchenyi István Egyetem), Zoltán Horváth (Széchenyi István Egyetem)

Matyas_Constans.pdf

Matyas_Constans.pptx

16:30 Deep particle tracking for pCT detector system

Hadron therapy is a form of cancer therapy, where we aim to destroy those cancerous cells that are hard to reach with surgery. Since this kind of approach differs from the classical gamma radiation therapy, the tomography methods used for that are not sufficient enough for Hadron Therapy . Proton Computed Tomography (pCT) is developed to achieve more accurate results for this kind of treatment. During pCT a detector system measures incoming particles that are scattered on a phantom. To track these particle we introduce 2 machine learning based solution. One of them is specifically focus on pCT detector system and the other one is a more general approach that we might use in other matching problems.

Speaker: Dudás Bence (Eötvös Loránd University)

GPU Day 2025.pptx

17:00 Image reconstruction with proton computed tomography

One of the most successful treatments in cancer therapy is proton therapy, with radiation planning being a key element. Photon CT is commonly used for this purpose; however, it does not provide sufficiently accurate information about the range of protons. Therefore, proton CT imaging is more favorable for radiation planning. Due to the Coulomb scattering of protons, it is important to calculate the Relative Stopping Power at the voxel level (thus, appropriate handling of trajectories is also required), for which several algorithms have been developed. The aim of my research is to test, further develop, and optimize a software package using the Richardson-Lucy algorithm developed in the Bergen Proton-CT Collaboration.

The simulations necessary for the research were performed using the Geant4 and Gate software. I optimized the framework using the Richardson-Lucy algorithm with appropriate methods for faster and more efficient operation. I verified the operation of the algorithm and image reconstruction on phantoms developed to measure the performance of medical imaging systems at different energies.

During my work, I managed to optimize the algorithm, significantly reducing the runtime. Based on the evaluation of phantom reconstruction, I found that the algorithm operates with the desired accuracy.

Among my long-term goals are further optimization and achieving clinical usability (including further reducing runtime).

Speaker: Zsófia Jólesz

GPUDay_JZs.pdf

17:30 Exponential distillation of dominant eigenproperties

Precisely estimating eigenstate properties of quantum many-body systems is a task of fundamental importance that is naturally demanding when approached using classical computation alone. While the incorporation of quantum computation provides promising alternative paths, their practical utility in near-term and early fault-tolerant quantum devices requires offloading as much computational burden to classical computation as possible. In this work, we propose a hybrid quantum-classical algorithm that extracts properties of a target eigenstate, which requires only a single quantum register with one ancilla qubit from the quantum device, and achieves an exponential suppression of erroneous contributions via classical post-processing. Combining insights from virtual distillation and random time evolution, the algorithm can target both ground and excited state properties. Moreover, it admits a flexible choice of techniques in both the quantum and classical algorithmic components. We prove rigorous performance guarantees and present optimized quantum circuit constructions. Through an extensive set of numerical simulations, we also demonstrate the applicability and scalability of our framework in both near-term and early fault-tolerant settings. Furthermore, we showcase its potential when combined with state-of-the-art classical simulation techniques using tensor networks.

Speaker: Bence Bakó (Wigner RCP)

gsprep_pres.pdf

Friday, 23 May

09:00 → 10:30 Session V

09:00 pLMSAV: A Delta-Embedding Approach for Predicting Pathogenic Single Amino Acid Variants

Predicting whether single amino acid variants (SAVs) in proteins lead to pathogenic outcomes is a critical challenge in molecular biology and precision medicine. Experimental determination of the effects of all possible mutations or those observed in pathogenic individuals is infeasible. While existing state-of-the-art tools such as AlphaMissense show promise, their performance remains insufficient for diagnostic applications, they are often challenging to run locally, and most are restricted to human sequences. To address these limitations, we developed pLMSAV, a simple yet effective predictor leveraging protein language models (pLMs). Our method computes delta-embeddings by subtracting the embedding of the mutant sequence from that of the wild type sequence. These delta-embedding vectors serve as input for a convolutional neural network used for training and prediction. To prevent data leakage, we trained our model on a well-characterized, labeled set of Eff10k and evaluated it on a non-homologous subset of ClinVar data. Our results demonstrate that this approach performs exceptionally well on Eff10k test folds and reasonably on ClinVar test sets. Notably, pLMSAV excels in resolving ambiguous predictions by AlphaMissense, also outperforming REVEL predictions of these cases. Therefore, we will integrat these REVEL-enhanced predictions into our widely used AlphaMissense web application (URL). We anticipate that incorporating delta-embeddings into other mutation effect predictors or mutant structure prediction methods will further enhance their accuracy and utility in diverse biological contexts.

Speaker: Dr Orsolya Gereben (Institute of Biophysics and Radiation Biology, Semmelweis University)

alphamissense-gpuday.pptx

09:30 AI applications from Astrophysics to Robotics

I will present AI applications in problems from Astrophysics to Robotics. Some of these developments and advancements are done within the Artificial Intelligence National Laboratory (MILAB), a state-of-the-art research facility focused on AI technologies.

Speaker: Vera Könyves (SZTAKI (Institute for Computer Science and Control))

veraKONYVES_GPUday_2025.pdf

10:00 Neural Networks in Physics Research

In this talk, we present three use cases demonstrating how AI tools are advancing research at the HUN-REN Wigner Research Centre for Physics as part of the AI 4 Science program.

1. Literature Review with Large Language Models (LLMs):
We leverage large language models to streamline literature searches and process complex scientific documents. Recent state-of-the-art open-source models are integrated into the HUN-REN GenAI 4 Science platform, and lighter-weight models can also be run efficiently on local hardware.
2. Generative AI for Scientific Writing and Coding:
We utilize generative AI to enhance scientific writing and assist code generation via tools like GitHub Copilot. This significantly accelerates the development of test cases, automatic documentation, and visualization workflows.
3. Exploring Physics-Informed Neural Networks (PINNs):
To deepen our understanding of physics-informed neural networks, we've launched a journal club focused on their theoretical and practical aspects. PINNs combine data-driven learning with the rigor of physical laws: they fit observed data while ensuring compliance with (possibly incomplete) differential equations and boundary conditions through automatic differentiation. This emerging field raises practical challenges, such as tuning network depth, width, activation functions, and loss balance; moreover, fundamental questions about whether a single network can represent the solution space for all initial conditions. While our exploration is still in the early stages, there is growing interest in applications ranging from tokamak plasma diagnostics to reconstructing electric potentials in spiking biological neural networks.

Speaker: Marcell Stippinger (HUN-REN Wigner FK RMI Komputációs Tudományok Osztálya)

Stippinger_NeuralNetsPhysics_GPUday.pdf

10:30 → 11:00 Coffee Break

11:00 → 12:30 Session VI

11:00 HUN-REN Cloud's new GPU resource management methodology in the AI4Science program

HUN-REN Cloud's primary goal is to support the Hungarian scientific community by providing essential e-infrastructure for their research. The HUN-REN Cloud is one of the main infrastructure pillars of the AI4Science program, which aims to enhance the use of artificial intelligence within the HUN-REN Research Network. This new program has opened up even more opportunities for researchers on the HUN-REN Cloud to accelerate their research utilising AI technologies which is necessitating a transformation of the current resource management system. A newly developed and implemented GPU resource management system is introduced to provide a more demand-responsive (differentiated) solution that makes cloud services available to more users in a more personalised way.

The HUN-REN Cloud is extended with a centrally managed, SLURM scheduler-based batch execution service, which aims to allocate GPU resources more dynamically. The SLURM services provide Singularity containers for a customisable execution environment and a Jupyter-based interactive mode for user convenience. In response to the growing interest in large language models (LLM), HUN-REN Cloud has introduced a PaaS service called GenAI4Science, which provides access to open-source LLMs. This service features a user-friendly web interface and OpenAI-compatible API access for LLMs deployed in a distributed environment with streamlined data management.

This talk will describe the new GPU resource allocation system, along with the new SLURM and GenAI4Science service, to support the HUN-REN AI4Science program.

Speaker: Attila Farkas (HUN-REN SZTAKI)

20250523_HUN_REN_Cloud_GPU_Day_2025_FA.pdf

11:30 Causality methods in time series analysis

Building models for multi-channel data can be complex, as it is often unclear which channels influence others and what the optimal time lag is. Allowing too many free parameters can be inefficient, requiring large amounts of data for reliable estimation and often leading to a less stable system compared to a model with fewer parameters. This challenge can be addressed by uncovering causal relationships. In this talk, we present a simple yet effective method for performing such an analysis.

Speaker: Antal Jakovac (Wigner RCP, Department of Computational Sciences)

GPU_days_2025.pdf

12:00 WebGPU - Compute kernels in your browser

WebGPU is a new, upcoming API to harness the power of GPUs. While there are already many competing standards on the market, WebGPU fills a particular gap in the spectrum by enabling applications to launch general compute kernels besides usual vertex and fragment shaders right in your browser. While this already opens up many possibilities for interesting applications, the fact that WebGPU is not exclusive to the web, but can also be used to build native desktop applications further improves its unique value proposition.
In this talk we review the main functionalities of the API, compare it to other similar technologies, discuss the current state of development and show some demonstration programs.

Speaker: Dániel Berényi (Freelancer)

dberenyi_WebGPU.pdf

12:30 → 14:00 Lunch break

14:00 → 15:30 Session VII

14:00 99999 Informatika

Speaker: Valentin Czuczumanov (99999 Informatika Kft.)

05232025_GPUDAY2025_99999.pdf

14:30 Cosmological N-body Simulations of the Rotating Universe

Everything in the Universe rotates. Expanding this principle to the Universe as a whole may seem like a natural step, yet the idea remains unexplored and is mostly dismissed as a curiosity. Recently, however, interest in this topic has gained traction, notably due to the work of Szigeti et el. (2025). In the era of precision cosmology, addressing the growing tensions within the currently accepted concordance model has become a priority, making the study of such new frontiers increasingly relevant. While a theoretical framework for a rotating Universe is already present, corresponding computer simulations are still virtually absent. Here, we present a numerical scheme for conducting and analyzing fully three-dimensional N-body simulations of a rotating Universe, implemented within the Stereographically Projected Cosmological Simulations (StePS) simulation code. We further compare results obtained using both spherical and cylindrical geometries.

Speaker: Balázs Pál (Wigner Research Centre for Physics)

pal-balazs-rotating-universe.pdf

15:00 The effect of extra dimensions on astrophysical observables

Kaluza and Klein proposed a theory with a compactified extra dimension, which may appear in high-energy phenomena, such as nuclear reactions, strong gravitational effects, or in the presence of superdense matter. In this work, I show how astrophysical observables will be modified in the presence of extra compactified dimensions.

The interior of a compact star is modelled as a multidimensional interacting degenerate Fermi gas, embedded in a static, spherically symmetric spacetime with extra compactified spatial dimensions. The equation of state of this extreme medium is given and compared to the standard models of superdense matter. The modification of the mass-radius relation of compact stars is calculated and compared to realistic star models and astrophysical observation data. The interaction strength has been determined for this extraordinary matter. Constraints on the size of the extra dimension have been estimated based on pulsar measurements [1-3].

[1] A. Horváth, E. Forgács-Dajka, G.G. Barnaföldi: "Application of Kaluza-Klein Theory in Modeling Compact Stars: Exploring Extra Dimensions", MNRAS, https://doi.org/10.1093/mnras/stae2637
[2] A. Horváth, E. Forgács-Dajka, G.G. Barnaföldi: "The effect of multiple extra dimensions on the maximal mass of compact stars in Kaluza-Klein space-time", IJMPA, https://doi.org/10.1142/S0217751X25420047
[3] A. Horváth, E. Forgács-Dajka, G.G. Barnaföldi: ”Speed of sound in Kaluza-Klein Fermi gas”, Accepted into: Acta Physica Polonica, DOI: 10.48550/arXiv.2502.04974

Speaker: Anna Horváth

AnnaHorvath_GPUDay2025.pdf

15:30 → 16:00 Coffee Break

16:00 → 18:00 Session VIII

16:00 The Importance of Randomness in ML

Everybody knows that machine learning models are initialized with random numbers. If we want to make our experiments reproducible, we have to use a random seed. But when and where should we use random seeds to freeze the entire environment? Sometimes, randomness occurs even when a seed is used.

I present real examples demonstrating the importance of randomness at different levels in machine learning, from model building and data analysis to training. Most researchers use a fixed random seed during experiments, but if they work with non-deterministic operations, the results may not be reproducible. Some pitfalls can be avoided, while others cannot, depending on the nature of the experiment and the structure of the model itself. However, understanding the real limitations of an experiment is crucial. In laboratory and production environments, this knowledge can save significant amounts of time and money in training models.

In the lecture, I present real code snippets to demonstrate how model training can fail due to a lack of randomness, such as when working with an unshuffled dataset. Researchers often rely on the default initialization methods for layers and nodes, which may yield predictable results but not necessarily the most optimal ones. This area is vast, even in classical machine learning. However, in the lecture, I will also showcase some use cases from the field of quantum machine learning

Speaker: Richárd Ádám Dr. Vécsey (szabadúszó)

DrVécseyRichárdÁdám_GPU_DAY.pdf

16:30 Laser ion acceleration using gold nanorods

This work investigates how integrating gold nanorods into laser targets enhances laser-driven ion acceleration. By exploiting the localized surface plasmon resonance (LSPR) of gold nanorods, we improve the coupling of femtosecond Ti:Sapphire laser pulses to the target. Numerical simulations reveal that resonant plasmonic excitations in the nanorods substantially intensify local electromagnetic fields and field gradients, concentrating laser energy near the nanostructures. This enhanced energy deposition increases the maximum ion energies compared to conventional flat targets, enabling more efficient ion acceleration within the preplasma region. We analyze key mechanisms, including Coulomb explosion and plasmonic ponderomotive acceleration, and demonstrate that tailoring nanoparticle geometry and arrangement is critical for optimizing near-field enhancement. These results present a promising route toward more compact, efficient ion sources and support future advances in laser-driven fusion.

Speaker: Christopher Grayson (Wigner Research Centre for Physics)

PP2025 - Ion Acceleration from plasmics.pdf

PP2025 - Ion Acceleration from plasmics.pptx

17:00 The use of the cuFFTDx library for performance optimization of Fourier-transform based GPU algorithms

Many GPU accelerated applications rely on the cuFFT library for fast and efficient Fourier transform implementations, however for certain algorithms it can be a performance limiting factor due to its strictly host-side API. Library functions cannot be called from code running on the GPU, hence unnecessary kernel launches, and host-device communication can occur when custom operations are needed to be performed before and after the transform. One possible solution to this problem is NVIDIA's cuFFT Device Extensions library (cuFFTDx) that provides FFT implementations easily integrable into GPU kernel code. This talk aims to present the main use-cases, the internal workings, capabilities and potential drawbacks of cuFFTDx alongside the C++ metaprogramming techniques behind its API. The practical use of the library will be demonstrated with the custom implementation of Welch's modified periodogram algorithm for spectral density estimation, highlighting the key differences between the standard cuFFT and the cuFFTDx approaches. In conclusion, the performance comparison of the two implementations will be presented based on runtime measurements.

Speaker: Bálint Tóth (University of Pannonia)

TóthBálint_GPUDay2025.pdf