Second-Gen European CPU for HPC
SiPearl, a processor designer supported by the European Processor Initiative, is about to begin shipments of its very first Rhea processor for high-performance computing workloads. However the firm is already engaged on its successor presently often known as Rhea-2, which is ready to reach generally in 2026 in Exascale supercomputers.
SiPearl’s Rhea-1 datacenter-grade system-on-chip packs 72 off-the-shelf Arm Neoverse V1 cores designed for HPC and linked utilizing a mesh community. The CPU has an hybrid reminiscence subsystem that helps each HBM2E and DDR5 reminiscence to get each excessive reminiscence bandwidth and first rate reminiscence capability in addition to helps PCIe interconnects with the CXL protocol on prime. The CPU was designed by a contract chip designer and is made by TSMC on its N6 (6 nm-class) course of expertise.
The unique Rhea is to a big diploma a product aimed to show that SiPearl, a European firm, can ship a datacenter-grade processor. This CPU now powers Jupiter, Europe’s first exascale sysatem that makes use of nodes powered by 4 Rhea CPUs and NVIDIA’s H200 AI and HPC GPUs. Provided that Rhea is SiPearl’s first processor, the challenge will be thought of as fruitful.
With its 2nd technology Rhea processors, SiPearl must develop one thing that’s significantly extra aggressive. That is maybe why Rhea-2 will use a dual-chiplet implementation. Such a design will allow SiPearl to pack extra processing cores and subsequently provide greater efficiency. After all, it stays to be seen what number of cores SiPearl plans to combine into Rhea 2, however a minimum of the CPU firm is ready to undertake the identical design methodologies as AMD and Intel.
Given the timing for SiPearl’s Rhea 2 and the corporate’s pure with to protect software program compatibility with Rhea 1, it’s cheap to anticipate the processor to undertake Arm’s Neoverse V3 cores for its second processor. Arm’s Neoverse V3 provide fairly a major uplift in comparison with Neoverse V2 (and V1) and may scale to as much as 128 cores per socket, which needs to be fairly first rate for HPC purposes in 2025 – 2026.
Whereas SiPearl will proceed creating CPUs, it stays to be seen whether or not EPI will handle to ship AI and HPC accelerators which are aggressive in opposition to these from NVIDIA, AMD, and Intel.
