The seventh fastest supercomputer in the world, Leonardo, is getting a major upgrade to take on AI workloads. The EuroHPC JU is spending €28 million to upgrade Leonardo to include new GPUs, CPUs and “high-bandwidth memory which will improve Leonardo’s performance by making it more efficient and precise in executing AI tasks,” the organization said.

The request for proposals page, posted on Thursday, is inviting companies to bid on implementing the upgrades.

The upgrade program, called LISA (Leonardo Improved Supercomputing Architecture) involves adding a new partition that will handle AI workloads.

Leonardo is classified as a pre-exascale computer, and delivers 241 petaflops of performance, according to the Top500 website.

The system, installed in Italy at Bologna Technopole, has Intel’s Xeon CPUs and Nvidia’s A100 GPUs. The system was made by Eviden/Atos.

It is hard to calculate the performance boost Leonardo will get as a lot depends on the relative power efficiency of the systems. Leonardo currently ranks 28th on the Green500 list, and energy efficiency is a priority for EuroHPC with the new AI partition.

A technical specification document outlines the upgrades to the system.

Leonardo could get LINPACK performance boosts from 45 to 65 petaflops in the new partition. The mixed precision performance boost or HPL-MxP could range from 300 petaflops to 650 petaflops.

Leonardo ranked fifth in the Top500 HPL-MxP ranking in June 2024, delivering 1.842 exaflops of performance. An additional 650 petaflops of HPL-MxP could put it in third place beyond Lumi, which had 2.35 exaflops of HPL-MxP performance. Typically, real-world performance doesn’t match up to theoretical performance.

EuroHPC’s Lumi and Leonardo weren’t designed for AI, and are far behind second-placed Frontier, which has 10 petaflops in HPL-MxP performance.

EuroHPC also wants the new partition to be based on the x86 architecture and have at least 165 nodes. Each node needs a minimum of 2 CPUs and at least 8 GPUs that are capable of training. Each node must have a minimum of 1TB of memory.

Specifically on the GPU, the document mentions the GPU must be able to train models, and “provide memory sharing with all the other GPUs installed in the same node.”

Also, the GPU’s HBM memory “must provide at least 80 GBytes.” That qualifies the latest GPUs from AMD and Nvidia.

The system is already built around Nvidia’s network cards, GPUs, and runs CUDA libraries. The memory sharing and software requirements tilt the requirements in Nvidia’s favor.

Each of the GPUs will connect to a network interface card, with a total interface connect speed of 3.2 terabits per second. The storage requirements include SSDs with fast IOPS. Data typically moves quickly on and off storage and memory, which the configuration will support.

EuroHPC JU expects the delivery of the system by April next year, with installation complete by July.

EuroHPC is adopting modular designs which makes the supercomputers easier to upgrade. For example, the supercomputers in Europe are being upgraded with quantum partitions.

EuroHPC’s first exascale supercomputer, Jupiter, is now being installed and is being designed for AI. The second exascale system called Alice Recoque, was announced earlier this year.