Europe's First Exascale Machine Gets Its European Chip

SiPearl delivered the first samples of its Rhea1 processor to Bull on 22 September 2026, and Bull has started fitting them onto the BullSequana XH3000 platform that will form JUPITER's CPU cluster, the two companies said in a joint statement issued that day. It is the first time a European-designed server processor will run inside a European-built supercomputer.

The chip is a large one by any measure. Rhea1 carries 80 Arm Neoverse V1 cores in a single package, each with two 256-bit Scalable Vector Extension units, alongside four stacks of high-bandwidth memory and four DDR5 interfaces, and totals 61 billion transistors. The Register, which reported the delivery on 23 September, puts the HBM2e at 64 GB per processor delivering 1.8 TB/s of bandwidth. That ratio of memory bandwidth to arithmetic throughput is the point: it suits the scientific codes that run badly on GPUs, the same reasoning that took Fujitsu's A64FX and Japan's Fugaku to the top of the Top500 in 2020.

Where Rhea1 actually sits

JUPITER, procured by the EuroHPC Joint Undertaking and hosted at the Jülich Supercomputing Centre at Forschungszentrum Jülich in Germany, became the European Union's first exascale system when it came online in 2025 and currently ranks fifth on the Top500 list. It did not get there on European silicon. The Booster partition that set the record runs on 24,000 Nvidia GH200 superchips, each pairing a Grace CPU with an H200 GPU. Rhea1 goes into the Cluster Module beside it, a CPU-only partition that The Register estimates at 1,300 nodes, 2,600 processors and roughly 5 petaFLOPS when complete.

Five petaFLOPS is a rounding error next to an exascale Booster, and the honest reading of this delivery is that its value is industrial rather than computational. Europe now has a domestic server CPU that has cleared design, fabrication and first integration into a flagship machine, with a named customer paying for it. SiPearl says the JUPITER equipment is its first Rhea1 order and its first significant revenue. The company employs 200 people in France, Spain and Italy, and its successor part, Rhea2, is slated for the CPU partition of Alice Recoque, the planned French exascale system, whose GPU side will use AMD MI430X accelerators.

The generation gap nobody is denying

Rhea1 is built on Arm's Neoverse V1 cores, a generation behind the V2 cores in the Grace CPUs already installed in JUPITER's own Booster, a consequence of a design cycle that has run for several years. SiPearl says benchmarking at its laboratories is delivering very encouraging results and that performance is in some cases exceeding expectations, but those are vendor statements about unconsolidated figures, and the company says the numbers will be published this autumn. Until an independent run appears, there is no way to judge how the part compares with a Grace or an A64FX on the workloads Jülich cares about.

SiPearl also makes a security claim that Nvidia and AMD do not, describing Rhea1 as free of back doors and kill switches. That is the argument the European Processor Initiative was funded to make good on, and it is the reason a modest CPU partition carries political weight that its petaFLOPS do not. Anders Dam Jensen, executive director of EuroHPC JU, said the delivery shows a European processor moving from innovation to deployment.

The comparison worth holding onto is with CXMT, which as Parallax Nexus reported on 20 September put an 11.95-nanometre DRAM platform into mass production using quadruple patterning because export controls keep extreme ultraviolet tools out of China. Both are sovereignty projects that accept a node or a core generation of disadvantage in exchange for a supply chain their government controls. Neither has yet shown that the penalty stays small as the frontier moves.