HPE – working with France’s HPC agency, GENCI, and its National Computing Center for Higher Education, CINES – announced a stellar win today at SC21: it will build France’s 70 peak petaflops Adastra supercomputer, scheduling another leading system for delivery in Europe on the heels of a string of EuroHPC system debuts. The system, slated for delivery and operation next year, will be used to support French academic and industrial research.

Adastra – a partial conjoining of the famous phrase “ad astra per aspera” (Latin for “to the stars through hardship”) – will consist of two complementary partitions. The first partition, scheduled to be operational by spring 2022, will consist of manycore scalar nodes based on next-gen AMD Epyc Genoa CPUs and 700GB of DDR5 memory per node, all networked with HPE Slingshot 11 (200Gbps). The second partition’s nodes, slated for operation by Q4 2022, will complement a third-gen AMD Epyc CPU with 256GB of DDR4 memory and quadruple AMD Instinct MI250X OAM GPUs, again networked by HPE Slingshot 11 (200Gbps). Adastra, which will be based at CINES in Montpellier, France, will have access to “expanded storage” to support complex modeling, simulation and AI workloads.

“This new Adastra supercomputer is a double challenging for CINES: massively switch users to boosting GPU while overcoming the electrical consumption,” said Boris Dintrans, director of CINES. “The 21x jump in computing power is dizzying compared to the current machine [Ed. note: HPE says “20X”], while power consumption only grows by 1.5x for green HPC. The GPU support provided by AMD, using the HPE Cray EX system, allowing Adastra to ramp up, is the key element that will ensure the success takeoff of researchers to the stars!”

CINES’ existing supercomputer, the 3.5 peak petaflops Occigen system, first posted to the Top500 in late 2014. Occigen, built by Europe’s own Atos, consists of three parts: a 2,106-node part with dual Intel Haswell CPUs and either 64GB of memory (1,053 nodes) or 128GB (1,053 nodes) per node; a 1,260 node part with dual Intel Broadwell CPUs and 64GB of memory per node; and a single-node large-memory part octuple Intel Haswell CPUs and 3TB of memory. (Occigen also has a visualization section with four dual-Broadwell nodes and 256GB of memory per node.)

It’s interesting to note that HPE’s system win — particularly for a successor to an Atos-built system — comes at a time when the EuroHPC Joint Undertaking and its members are working to reconcile two occasionally competing priorities: the desire for the best, most cost-effective system and the desire to develop European HPC sovereignty.

Beyond the vendor, Adastra also represents another win for AMD in its increasingly successful siege against Intel’s market dominance (AMD powers all four new systems in the top 15 of today’s Top500 announcement). Even more specifically, it’s a win for AMD’s newly announced line of remarkably powerful multichip MI200 GPUs, each of which delivers 47.9 peak double-precision teraflops.

While Adastra is being deployed to generally support French research, the organizations are highlighting a few key application areas of particular interest: the advancement of renewable energy, aimed at improving France’s energy security and reducing its carbon footprint; materials research for next-generation battery technology; and improved drug design and treatment research, which resonates in a quasi-post-pandemic world for obvious reasons.

“Adastra will allow French research teams to position themselves even more strongly on the path to exascale, and to prepare for the change associated with extremely large-scale resources and new technologies/services,” said Philippe Lavocat, chairman and CEO of GENCI. “The realization of this major step came from the dialogue and the expertise of CINES, GENCI and HPE teams.”