Programmable accelerators boost supercomputer speeds
Many state-of-the-art consumer and industrial services, including aeronautics, automobile design, and climate models, depend on emerging high-performance computing (HPC) applications. Meeting the rising demand for such services requires delivering more powerful computers at lower prices. For some HPC applications, field-programmable gate arrays (FPGAs) are a promising solution. These accelerators can be programmed after manufacturing to suit tailor-made applications. FPGAs can greatly outperform competitive processors such as CPUs or GPUs, granting these significant advantages for HPC applications. “FPGAs excel at parallel and specialised computations,” explains Iakovos Mavroidis, researcher at the Technical University of Crete and OPTIMA project coordinator. “They are particularly effective for parallel, irregular or low-latency tasks in HPC, such as cryptographic algorithms, signal processing and scientific simulations,” he says. In the OPTIMA project, Mavroidis and his colleagues used HPC systems integrating FPGAs to develop a series of open-source libraries and industrial applications, such as fluid flow models, with significantly higher energy efficiency. “Developing use cases on OPTIMA FPGA-based infrastructures was challenging, as we had to balance flexibility and manage resources effectively to ensure good performance without overloading hardware,” adds Mavroidis. “However, careful planning, thorough testing, and teamwork among FPGA design, hardware and software teams helped us tackle these challenges.”
Integrating FPGA-based accelerators
The OPTIMA team – a consortium of partners from across Europe – started by understanding the requirements of industrial applications and determining the necessary features for the eventual OPTIMA open-source libraries. Next, the researchers deployed two hardware platforms: one which interconnected four FPGA-based accelerators ; the other based on Maxeler’s dataflow engines and programming model. The team then started coding on the FPGA-based platforms, thoroughly testing each aspect, while continuously adding optimisations to the system to make it more efficient. The researchers documented their actions and shared their work with the open-source community.
Demonstrating prototypes
OPTIMA successfully developed four HPC applications on two FPGA-based prototype accelerators. These applications include tools for underground analysis, powerful ‘MESHFREE’ simulation tools and AI-based robotics simulations – all applications which deal with huge amounts of data and complex tasks. Additionally, the project created an open-source library on FPGAs. This consists of 31 hardware components which support fundamental linear algebraic operations and computer-aided engineering problem-solving methods, which are crucial for artificial and machine learning applications. “OPTIMA demonstrated how using FPGA-based technologies can improve HPC systems for industry,” says Mavroidis. As the library has been made available to everyone, developers can easily move applications and legacy code to FPGA-supported HPC systems. The project was carried out with support from the European High Performance Computing Joint Undertaking (EuroHPC JU), an initiative set up to develop a world-class supercomputing ecosystem in Europe. “The OPTIMA project supports EuroHPC JU’s goal of making Europe a leader in HPC, and it’s a great example of a small business-driven EuroHPC project,” notes Mavroidis.
Business acceleration
The team expects the OPTIMA open-source library to continue to grow, to become a valuable resource for software developers aiming to optimise their applications for FPGA-supported HPC platforms. The expertise gained from adapting applications to the OPTIMA platforms is now being applied by partner SMEs. This includes developing new applications for advanced cloud systems, AI accelerators, GPUs and other chip designs.
Keywords
OPTIMA, EuroHPC JU, FPGA-based accelerators, prototypes, HPC, industry, business, hardware, optimise, supercomputer
