Fraunhofer ITWM at ISC

High Performance Computing and Artificial Intelligence

Press Release / May 08, 2024

The division »High Performance Computing« of the Fraunhofer Institute for Industrial Mathematics ITWM, known for its exascale-capable programming tools and the BeeGFS parallel file system, is coming to Hamburg with innovative solutions for the HPC community in the fields of parallel programming and Artificial Intelligence. At the joint booth with ThinkparQ, a Fraunhofer ITWM spin-off, the experts will be presenting various innovations from the institute at ISC High Performance Computing 2024.

On site, the researchers will be demonstrating CARME, a new open source software environment. The aim is to »simplify interactive machine learning on HPC systems«. GaspiLS pushes the limits of scalability for iterative solvers, which are fundamental for CFD and FEM simulations. With the STX processor board, the Fraunhofer ITWM is presenting the first result of its HPC processor development. ThinkparQ will also be presenting its ambitious BeeGFS roadmap.

In terms of green computing, the institute encourages visitors to learn about the latest developments of the Fraunhofer ITWM HPC processor to accelerate simulation codes. The Neural Architecture Search Engine (NASE) demonstrates massive energy savings through optimization and hardware adaptation of DNNs.

Networking at Booth L40

The team from the mathematical institute invites all visitors from the ISC 2024 to visit the joint booth and find out about the latest technologies. Intensive discussions and new networking opportunities will provide an opportunity for exchange.

A highlight awaits visitors to the Fraunhofer ITWM stand: The talented and up-and-coming artist, Gesche Karnik, will be on site and will be creating new works during the fair. Here, interested visitors will have the opportunity to experience the creation of the artworks live and to talk to the artist in person.

© iStockphoto / Fraunhofer ITWM
Carme Deep Learning

CARME – Softwarestack for Interactive ML on HPC Systems

We are delighted to introduce our open-source softwarestack, CARME, at the International SupercomputingConference (ISC) 2024. CARME encapsulates our expertisein both Artificial Intelligence (AI) and High PerformanceComputing (HPC), and can be effortlessly integrated into yourHPC systems. With CARME, you can concentrate fully onyour AI development objectives without needing any priorexperience or knowledge of workload management tools orhardware. That's because CARME handles all of this for you.

CARME is designed to be user-friendly, secure, and flexible. Itprovides web-based access and two-factor authentication(2FA), incorporates container technology, and offers a varietyof other features. All these make it an easy-to-use and robusttool to manage both interactive and batch AI workloads onHPC systems. At ISC we present version 1.0 with its easyinstallation process.

More information about CARME can be found here.

GaspiLS –A Scalable Iterative Linear Solver for CFD- and FEM-Simulations

GaspiLS, our scalable iterative solver library, hasdemonstrated exceptional scalability in Independent SoftwareVendor (ISV) and industry codes. Given the prevalence ofworkstations with over 100 cores and small High PerformanceComputing (HPC) installations with thousands of cores, thestrong scalability of our solver library is indispensable forengineering solutions.

GaspiLS is not just a library, it's a toolbox that enables you todesign the optimal solver for your code. It features highlyscalable Algebraic Multigrid (AMG) and hybrid Incomplete LU(ILU) preconditioners that outperform competitors such asHYPRE and PETSc. In addition, GaspiLS is open source,fostering a collaborative approach to improving itscapabilities. GaspiLS can be integrated seamlessly into anycode that uses Message Passing Interface (MPI) or GlobalAddress Space Programming Interface (GPI), making it aversatile choice.

More information about GaspiLS can be found here.

© freepik / iStockphoto / Fraunhofer ITWM
GaspiLS Keyvisual

NASE – Neural Architecture Search Engine

The NASE (Neural Architecture Search Engine) at FraunhoferITWM is a powerful AI tool designed to optimize NeuralNetwork models towards the underlying hardware – e.g. AIaccelerators, customs AISCs and FPGAs. This serviceautomates the process of finding efficient AI models tailoredto specific performance needs such as speed, latency andpower consumption. By incorporating unique hardwarecharacteristics into model design, NASE ensures that eachneural architecture is both optimal and custom-fitted to theuser's requirements. This solution leverages advancedalgorithms and extensive computing resources to deliverready-to-use, efficient neural networks for real-worldapplications.

At our booth, we'll show how NASE enhances medicaldevices by optimizing Neural Networks for low energyconsumption, critical in wearables like ECG monitors. Thisoptimization significantly reduces energy usage, achievingdramatic improvements across several orders of magnitude.

More information about NASE can be found here.

The STX Processor

This year, the Fraunhofer ITWM booth will feature the STXProcessor card, showcasing the PCIe 5 Board equipped withfour STX chiplets. These chiplets excel in bandwidth-limitedapplications, particularly with stencil algorithms.

The system offers developers a user-friendly programminginterface through a C++ and OpenMP layer, efficientlymasking the underlying complexity. Our compiler team hassuccessfully optimized this setup to deliver peak performanceacross various simulation kernels, streamlining developmentand enhancing computational efficiency.

More information about STX-Processors can be found here.