[A GREAT STORY] State-of-the-art computing infrastructure supporting research in artificial intelligence

Artificial intelligence is a valuable tool for many sectors, so much so that few now do without it. However, machine learning methods, such as machine learning and deep learning, require significant computing power. And this is true even before they are put to practical use: the training phase, during which the algorithms “learn” to solve the problem presented to them, requires an immense amount of computation.

This need has long existed at Télécom Paris, a member school of the Carnot TSN Institute recognized for its expertise in AI. And what once primarily concerned the Image, Data, and Signal (IDS) department of the Laboratory for Information Processing and Communication (LTCI) has now expanded to include all research teams.

More than 190 GPUs to train AI algorithms

In response to this growing demand, the school established, ten years ago, a computing infrastructure dedicated to AI-based research, with support from the Carnot TSN Institute. “The first machines arrived starting in 2016, and the platform has continued to expand every year since,” explains Nicolas Bouché, head of the Computing Resources Engineering Unit at Télécom Paris. “ It underwent a major transformation five years ago, expanding its hardware portfolio and shifting from management by the IT department to integration within the Research Division.”

Today, the infrastructure consists of 78 servers and more than 190 GPUs (Graphics Processing Units), which are cornerstones of AI model development. However, in the IT industry, graphics processing units were originally designed for displaying and creating images and videos. But their remarkable ability to quickly perform a large number of calculations has made them valuable—even indispensable—for AI, particularly during the training phase.

Thanks to this comprehensive list of state-of-the-art equipment, the Télécom Paris platform boasts a total computing power of 7.3 petaflops—or 7.3 million million million operations per second. This figure ranks it fourth among the best infrastructures available on a regional scale. “The support of the Carnot TSN Institute is essential to providing this level of service,” emphasizes Nicolas Bouché. “It helps fund the regular purchase of new machines, ensuring we can offer state-of-the-art equipment that keeps pace with the rapid pace of technological change, thereby ensuring our long-term sustainability. This support is all the more crucial as costs continue to rise, both for GPUs and for RAM, which is also essential.”

Optimized resource management based on demand

While Télécom Paris’s AI computing infrastructure is among the regional leaders in this field, it does not intend to compete with supercomputers on a national scale. “These are platforms with greater resources and a different set of goals,” notes Nicolas Bouché. “In fact, our approach complements theirs: we offer, in a way, a sandbox that makes it easy to test AI models. And if the test is successful, supercomputers are the best way to scale up, building on code that has already been proven. Furthermore, we emphasize close collaboration with our users, whom we support throughout their projects.”

To this end, the platform is managed by a team of four people, led by Nicolas Bouché, and is fully integrated into the research teams at Télécom Paris. And for good reason: it is these individuals—researchers, as well as doctoral students, postdoctoral researchers, and interns—who primarily benefit from the infrastructure. Students, however, can also access some of the equipment to carry out their projects.

Specifically, once users have written their algorithm code, they can request to run it on available resources, according to their desired configuration. “ They can choose one or more servers and the number of GPUs allocated to each,” explains Nicolas Bouché. “To manage and distribute requests optimally, we use a software tool called Slurm, which acts as a resource scheduler. In addition, the tool informs users about resource availability, helping them adjust their requested configuration based on their needs.”

Furthermore, rules have been put in place to prevent congestion and excessively long wait times. For example, a user cannot exceed a certain number of projects running simultaneously; a new, highly sought-after resource cannot be continuously accessed by the same project for more than one day, etc.

Local data storage and flexibility

For Télécom Paris research teams and students, the school’s infrastructure offers not only computing power but also control over their data. The data is stored locally in the school’s server room, located on the Saclay campus. This helps limit the risk of leaks of potentially sensitive or confidential data—a danger that is often present when using commercial platforms.

And compared to national supercomputers, Télécom Paris’s AI computing infrastructure offers greater flexibility. “When using a supercomputer, a user brings their own code and runs it on the available machines,” explains Nicolas Bouché. “ We offer a similar approach, but with greater flexibility: users can adjust their code in real time on the GPU, which provides agility that is often invaluable at this stage of algorithm development.”

In addition, it is possible to reserve certain parts of the infrastructure for private use, particularly for projects with external partners. “Sometimes, collaborative research projects are accompanied by private funding intended for the acquisition of specific equipment, ” says Nicolas Bouché. “This equipment is then integrated into the infrastructure but can be allocated to a private partition dedicated to the project in question. And at the end of the contract, depending on what has been agreed upon, the equipment may be transferred to one of the platform’s public partitions. ” In addition, the calculations performed using Télécom Paris’s resources and the results obtained can, of course, contribute to larger-scale projects.

From Télécom Paris’s computing infrastructure to that of IP Paris?

For example, the HI-AUDIO project (Hybrid and Interpretable Deep Neural Audio Machines) is among those that have benefited from the platform. Funded by a grant from the European Research Council, it aims to develop controllable, resource-efficient AI models applied to sound. These models could then be used for the transcription, transformation, or synthesis of audio signals, particularly for source separation, music generation, or the removal of reverberation effects in speech signals.

Currently, the computing infrastructure is reserved exclusively for Télécom Paris staff and students. However, the long-term goal is to make it available to all teams at the Institut Polytechnique de Paris (IP Paris). This transition is already underway, as the server room at the Saclay campus has, for several years now, housed machines belonging to ENSTA—another IP Paris school and a member of the Carnot TSN Institute. Although fewer in number, these machines are also managed using Slurm software, marking a first step toward a shared infrastructure for the entire IP Paris network.

Before this ambition becomes a reality, the infrastructure will continue to be expanded each year. “It’s essential to maintain a steady pace in improving the platform: this isn’t a sprint, but a marathon,” explains Nicolas Bouché. “In this regard, the support from the Carnot TSN Institute is crucial, as it allows us both to invest roughly the same amount each year and to have visibility into the coming years.”