itestro - Fotolia
France is building a supercomputer designed to handle AI workloads, as well as traditional high-performance computing applications.
It may well be the largest AI supercomputer in Europe, and one that points to future configurations in high-performance computing (HPC). The just announced Hewlett Packard Enterprise (HPE) system, named Jean Zay after a World War II French resistance hero, is expected to be put into production in October 2019.
The AI supercomputer will rely on a high number of graphics processing units (GPUs), as well as flash storage. The Jean Zay also gets France closer to becoming a leader in AI research and industrial development, a goal that the French government set last year.
"This supercomputer is vital," said Stéphane Requena, director of innovation at GENCI, the government-backed research organization that spearheads HPC computing in France, in an email. "It will target first more than 300 French researchers, developing and scaling out new AI models, and then more than 1,000 researchers using deep learning and machine learning."
Pulling back the curtain
The design for supercomputers is shifting rapidly to support mixed use applications, or converged systems, that support both traditional HPC and AI workloads. The Jean Zay will have 1,789 compute nodes with two sockets each with 20 cores to equal 71,560 Intel Xeon CPU cores. It will also have 261 nodes with four Nvidia Tesla V100 GPUs, which HPE said will total 1.33 million cores.
"The French supercomputer has a somewhat unique configuration with the extra GPUs, which gives more weight to it being dedicated to artificial intelligence, machine learning and deep learning" applications, said Dan Olds, HPC analyst and partner at consulting firm OrionX.
France's system will have four GPUs per node, which is double and sometimes more than double the GPUs found in a typical HPC system, according to Olds. Multi-GPU support is "the most obvious characteristic for a machine that is going to be utilized for AI," he said.
Stéphane Requenadirector of innovation, GENCI
GPUs, processing chips that grew out of the video game industry, use parallel processing power for computationally intensive tasks, which include AI-related applications.
The system will also use flash storage with read/write capability of more than 300 GB per second from storage provider DataDirect Networks.
The French AI supercomputer will have a theoretical peak performance of 14 petaflops, which would put it around the 20th most powerful supercomputer in the world, according to the most recent Top500 supercomputer listing.
As of November, the globe's leading supercomputer on the Top500 list is IBM's Summit, which has 9,200 IBM Power9 CPUs and 27,650 Nvidia Volta GPUs. This computer has a peak performance of 200 petaflops and is housed at Oak Ridge National Laboratory. IBM's Summit was also designed to handle AI-specific workloads.
A converged system
France will use this supercomputer to boost its science research and economic development. The AI supercomputer is optimized to process large data sets to solve HPC computing problems both within the research community, as well as the public sector, such as disaster preparedness.
The role for HPE goes beyond building the AI supercomputer, according to Pierre Hoffer, HPE director of Europe, Middle East and Africa presales for HPC and AI.
HPE has created an AI marketplace in France and has developed partnerships with independent software vendors that specialize in certain technologies and industries, such as automotive and banking. HPE currently has about 15 partners but plans to expand it, as well as open it up to other countries, Hoffer said.