high performance computing on graphics processing units: hgpu.org

The digital revolution of our society is driven by major technological advancements, enabled not only by the growing capabilities of computers but also by the evolution of their uses. These developments result from a complex interaction between what we can do, what we know how to do, and what we want to do, all within a constantly changing context. Computing, at its core, can be reduced to its ability to perform calculations, as processors do. It is all about computations and data. The performance of processors enables certain methods, which in turn generate new uses, creating demand for even more powerful processors to speed up these methods. The use of graphics processing units (GPUs) for deep learning is a striking example. Efficient utilization of increasingly powerful hardware has long been a pursued goal. High-performance computing (HPC) is a specific field of computing — both a research area and an essential tool for the development of digital applications. It is a vast domain with its own challenges, yet inherently linked to other sectors such as algorithms, mathematics, networking, and hardware. HPC research follows hardware advances. Each new innovation may necessitate the reimplementation of classic methods, either to adapt to the new features or to fully exploit their specifics. Methods enabling more general and flexible adaptation thus become crucial. In this domain, implementation plays a central role — not only to maximize program efficiency by fully utilizing hardware potential but also to structure tools and applications that may require years of development and rely on millions of lines of code. Sometimes, significant implementation work is necessary simply to assess the potential of a new approach. Document organization This document begins with a description of the current HPC ecosystem and the research and development processes for applications that depend on it. We will examine the main challenges in this field, including the layered software organization, which often complicates HPC application development, maintenance, and optimization. Another recurring issue is the lack of solution portability. Applications developed for a specific hardware architecture are not always easily adaptable to other environments, limiting their flexibility and longevity. The transfer cost can be eliminated or reduced when generating an optimized solution is performed automatically by a tool. However, a major challenge lies in the difficulty of producing generic tools. Current HPC applications are often rigid and designed to meet very specific needs, making their reuse or extension difficult without considerable adaptation work. Furthermore, the way these systems are developed often focuses on the detailed description of how to do — i.e., the technical procedures to achieve a goal — rather than focusing on clearly expressing what we want to do, the ultimate objective. This procedural approach constrains innovation and adaptability, especially in a rapidly evolving technological environment. After identifying these challenges, we will present the solutions we have implemented to overcome them. We will detail the strategies we have adopted to improve application portability, increase their generality, and facilitate the expression of intentions rather than procedures. This will include techniques aimed at simplifying development, reducing software layer complexity, and enabling greater flexibility in using hardware architectures. Finally, we will conclude by proposing a new prospective approach based on emerging trends in the HPC field. This approach will explore ways to rethink software and hardware architecture to better anticipate and meet future needs while allowing for improved scalability and modularity of applications. We will thus outline the foundations of a new vision of high-performance computing, more suited to the technological and human challenges of tomorrow.