SiPearl begins delivering Rhea1 CPUs to Bull for Jupiter's CPU partition

Jupiter, the EU's first exascale supercomputer, currently relies on 24,000 Nvidia GH200 superchips for its booster section, which holds the European performance record. The new development involves SiPearl delivering its Rhea1 CPUs to Bull, the system builder, for a separate CPU-only partition. This partition is distinct from the GPU-heavy booster and is intended for workloads that do not benefit from accelerated computing.

The Rhea1 is an 80-core processor based on Arm's Neoverse V1 architecture, featuring 61 billion transistors. It includes four stacks of HBM2e memory totaling 64 GB, providing 1.8 TB/s of bandwidth, alongside four DDR5 channels. This high memory-to-compute ratio is specifically tailored for HPC and scientific applications that require efficient access to large datasets, a characteristic that previously favored systems like Japan's Fugaku.

While the Rhea1's Neoverse V1 core design is older than the V2 cores used in the Grace CPUs within Jupiter's booster, its inclusion is significant for digital sovereignty. The CPU partition will eventually span 1,300 nodes with 2,600 processors, estimated to produce 5 petaFLOPS. Although this output is modest compared to the exascale booster, it provides essential capacity for non-GPU workloads within the same system.

The delivery signals a broader trend in European HPC infrastructure, as SiPearl's Rhea2 chips are also slated for the Alice Recoque supercomputer. This move allows the EU to diversify its hardware supply chain, ensuring that critical scientific computing capabilities are not solely dependent on US semiconductor manufacturers. The integration of domestic silicon into a top-five global supercomputer serves as a proof of concept for the region's sovereign computing ambitions.