Fresh off its inauguration just weeks ago, the EuroHPC-organized, Cineca-operated Leonardo supercomputer has been approved for a major expansion called LISA. The expansion is expected to increase Leonardo’s already formidable computing power by around 100 petaflops via two additional modules: one based on high-bandwidth memory nodes, another based on next-generation GPU nodes.
Leonardo, based in Italy, debuted on November’s Top500 list, taking the fourth spot with 174.7 Linpack petaflops using just around two thirds of one of its booster module. The Atos-built system, based on the “Modular Supercomputing Architecture” (MSA), includes two core modules: first, the aforementioned booster module, which features 3,456 nodes, each with an Intel Xeon “Ice Lake” CPU, quadruple custom Nvidia Ampere GPUs (64GB) and 512GB of memory, adding up to an aggregate 240 Linpack petaflops. Second, the data-centric module: 1,536 nodes, each with dual Intel “Sapphire Rapids” CPUs and 512GB of memory; this module delivers just around 9 Linpack petaflops. That partition is expected for installation around early 2023.
The new upgrade is called LISA, which ostensibly stands for “Leonardo Improved Supercomputing Architecture” – but which actually, of course, is mostly a nod to the Mona Lisa (in keeping with Leonardo’s namesake). LISA will add two new modules: first, a module with conventional nodes that utilize high-bandwidth memory, aimed at improving the performance of tasks requiring fast data transfers between the memory and the CPU. Second, what Cineca is calling a “high-end accelerated module” powered by “next-generation GPU server nodes.” Cineca adds that this second module will provide “top-of-the-market efficiency in terms of performance per watt.”
Cineca says that the two modules will constitute a considerable increase in performance as well as a “broadening of the use cases that can be well-addressed with the supercomputing system.” Cineca had previously submitted an expression of interest in performing the upgrade to the EuroHPC Joint Undertaking – the EU’s concerted supercomputing play – which recently approved the project.
Leonardo is the latest EuroHPC supercomputer to deploy, leaving only two from the initial slate of eight systems that was announced in June 2019. The first of the remaining systems is a petascale supercomputer called Deucalion, based on Arm and AMD CPUs as well as Nvidia GPUs, which is expected to launch any day now in Portugal. The second remaining system is MareNostrum 5, the last of the three pre-exascale systems, which is expected sometime in the back half of 2023 at the Barcelona Supercomputing Center. MareNostrum 5 will be powered by a range of Intel and Nvidia hardware. EuroHPC also announced its first exascale system, JUPITER, earlier this year, along with four sites for additional petascale or pre-exascale supercomputers – learn more about that here.