Thursday, September 24, 2026
AI 인프라 · 뉴스 & 분석
반도체·하드웨어리포트
반도체·하드웨어 · 리포트

JUPITER supercomputer receives SiPearl's Rhea1 CPUs, advancing Europe's domestic semiconductor capability for high-performance computing.

SiPearl Rhea1 deployment in JUPITER strengthens EU strategic autonomy in HPC/AI silicon, reducing dependence on US/non-EU suppliers.
업계 전문지Slicast · 2026년 9월 23일 18:32 UTC · 글로벌 · 출처: The Register
중요도 70

Europe's most powerful supercomputer, Jupiter, is receiving its first domestic silicon. On Tuesday, French chip designer SiPearl began delivery of its Rhea1 CPUs to Bull for integration into the system. Jupiter, which came online last year, became the European Union's first exascale supercomputer and currently ranks fifth on the Top500 list of publicly known supercomputers. However, that achievement rests on American technology—the Jupiter Booster section uses 24,000 GH200 superchips from Nvidia, combining a single Grace CPU with an H200 GPU. SiPearl's Rhea1 chips will power a second, CPU-only partition, marking a turning point for sovereign European supercomputing.

Rhea1 has been under development for several years. Based on Arm's Neoverse V1 core designs, the chip shows its relative age—the Grace CPUs powering the booster already use Arm's newer V2 cores. Still, Rhea1 remains significant. The 80-core chip features 61 billion transistors and was designed specifically for HPC and scientific workloads not well suited to GPUs, resembling the Fujitsu A64FX CPUs that powered Japan's Fugaku supercomputer to the top of the Top500 in early 2020.

Each Rhea1 processor includes four stacks of HBM2e memory totaling 64 GB with 1.8 TB/s of bandwidth, plus four DDR5 memory channels. Many HPC workloads benefit from this high ratio of memory bandwidth to FLOPS—the characteristic that explains Fugaku's strong performance on HPCG benchmarks relative to HPL rankings.

"Its high memory bandwidth is particularly well suited to applications that benefit from efficient access to large amounts of data," said Thomas Lippert, director of the Jülich Supercomputing Centre. "Bringing this European processor technology into Jupiter is an important step in making these capabilities available to the scientific community."

Anders Dam Jensen, executive director of EuroHPC JU, called the delivery a major milestone for the EU. "This achievement demonstrates what Europe can accomplish by bringing together research, industry and world-class infrastructure to build a stronger European HPC ecosystem," he said.

When complete, Jupiter's CPU partition will comprise 1,300 nodes totaling 2,600 processors with an estimated combined output of 5 petaFLOPS. While modest compared to the GPU booster, the CPU partition remains relevant since not all workloads can leverage GPU acceleration.

SiPearl's advance extends beyond Jupiter. The Alice Recoque supercomputer is slated to use the company's Rhea2 chips in its CPU partition and AMD's MI430X accelerators in its GPU partition.

원문 보기
JUPITER supercomputer receives SiPearl's Rhea1… · Slicast