Textarossa: preparing for the high-performance computing revolution

HPC - central to many challenges facing society

Enhanced by artificial intelligence (AI) and data analysis, high-performance computing (HPC) is behind some of the most important subjects for our future, whether it’s anticipating the impact of climate change on our planet, exploring materials in order to create innovative new medicinal products and treatments, optimising the shape of wind turbine blades to boost output or designing next-generation materials which are more lightweight, hard-wearing and durable.

This technology is central to Textarossa (Towards EXtreme scale Technologies and Accelerators for euROhpc hw/Sw Supercomputing Applications for exascale), a major EU H2020 project that was launched on 1st April 2022. Inria is one of 11 partners (see inset) involved in the project, which has been allocated a budget of 6 million euros over 36 months and is being coordinated by the Italian National Agency for New Technologies, Energy and Sustainable Development (ENEA).

The supercomputer revolution

Bérenger Bramas, a young IT researcher, is responsible for Inria’s contribution to this EU project, working with the Camus project team, which is based in Strasbourg and which falls under the supervision of the Inria Nancy - Grand Est research centre and the ICube Laboratory. This team specialises in the parallelisation and optimisation of multicore processors, developing methods enabling them to be used effectively in HPC.

Bérenger Bramas broke down the thinking behind Textarossa: “The 2020s are expected to see the emergence of european exaflop supercomputers or exascale computing, capable of carrying out 10¹⁸ operations a second, or one billion billion operations a second. This processing power is extraordinary: just one of these supercomputers would be as powerful as the 100 most powerful supercomputers currently in use in Europe put together. The USA, China, Japan and the EU have already launched ambitious exascale projects. Textarossa has its sights set on preparing for this digital revolution by evaluating the potential of new solutions which could, eventually, be deployed in such machines.”

The challenge of energy frugality

A number of scientific obstacles will need to be overcome in order for these digital behemoths to be developed and used. Chief among these is improving their energy frugality: supercomputers might boast unrivalled processing power, but this comes at the cost of several million euros worth of electricity each year. They also generate an incredible amount of heat, requiring powerful cooling systems.

As a result, some serious innovation will be needed in order to build and operate exascale computers. The Textarossa consortium (the name of which conjures up speed, in reference to a famous Italian sports car) has set itself a number of objectives: to develop new, super-fast IT materials; to design innovative cooling equipment; and to develop next-generation algorithms and software. The project is expected to have applications not only in the traditional fields of HPC - such as digital simulation - but also in emerging fields such as high-performance artificial intelligence (HPC-AI) or high-performance data analysis (HPDA).

Improving digital and energy performance

Inria’s involvement in the project is centred around its software component, its aim being to develop a new processing unit for HPC, before then evaluating its digital and energy performance. “Supercomputers break tasks down into millions of sub-tasks, which are then carried out simultaneously by just one processor”, explains Bérenger Bramas. “Machines currently use two types: CPUs (central processing units, like the ones found in computers) and GPUs (graphic processing units, used widely in the video game industry). Our focus will be on the potential of new FPGAs, circuits which can be programmed on request, making them particularly well-suited to HPC algorithms.” 

How are tasks to be sequenced by CPUs, GPUs and FPGAs within a supercomputer? How can you get the most out of their individual power capacities? What performance levels can be expected of FPGAs and how can these be improved? What criteria should be used to evaluate gains in terms of digital and energy efficiency? The research carried out by Inria and its partners should provide tangible answers to these questions, which will be of great interest to the entire HPC community.

Innovative project management

Bérenger Bramas will be working in close collaboration with his colleagues from the University of Bordeaux Inria Centre (Olivier Beaumont, Lionel Eyraud-Dubois, Brice Goglin, Abdou Guermouche, Raymond Namyst and Samuel Thibault), focusing on the key issue of sequencing. Their aim will be to demonstrate the performance levels of FPGAs. They will also be exploring the algorithmic aspects of HPC, anticipating results which could be of interest to both mathematicians and computer scientists, especially in linear algebra (a branch of mathematics used in HPC), and FMM (a digital method popular with engineers which can be used to reduce simulation time, for calculating noise from air or land transport, for example).

“Textarossa will give us an opportunity to test out the potential of innovative new approaches to HPC”, explains Bérenger Bramas. “Without this step, we wouldn't be able to deploy any innovations on future exascale machines with any confidence”. The project is also innovative in terms of its coordination, which involves the use of a “codesign” process, each partner lending their own expertise or ideas, while adapting their developments to others’ results. What gives supercomputers such incredible processing power is the way in which they distribute tasks and effectively coordinate their execution. The researchers who make up the Textarossa consortium are well aware of this - and intend to apply the same principle in the hope of achieving the same level of efficiency, drawing on agile project management.