SiPearl will cooperate with AMD to build exascale supercomputing system
SiPearl is a chip design company based in France, mainly designing chips based on Arm architecture for supercomputers. Recently, SiPearl announced that it will cooperate with AMD to jointly build an exascale supercomputer. Under the terms of the deal, the new supercomputing system will use SiPearl’s Arm-based Rhea SoC, as well as AMD’s Instinct line of computing cards.
The collaboration also includes building software compatible with the Rhea SoC and evaluating the interoperability of AMD’s ROCm with SiPearl’s Rhea SoC. ROCm is an open software platform for accelerated computing and does not limit the programming language, suitable for large-scale computing, and supports multi-GPU computing. Obviously, SiPearl also wants to use ROCm for its CPU. After all, HPC software for Arm architecture is still in its infancy.
Rhea SoC uses 72 cores based on the ARM Neoverse platform (codenamed Zeus), interconnected using a mesh network structure, 68 of which have their own independent L3 cache, and are equipped with some additional modules, manufactured using TSMC’s 6nm process. One of Rhea’s very characteristic designs is the hybrid memory system, including 4 sets of HBM2e memory controllers and 4 to 6 sets of conventional DDR5 memory controllers, combining the high bandwidth of HBM2e memory and the high capacity of DDR5 memory.
Although SiPearl promises that Rhea SoC has excellent performance, the performance is not enough to meet the computing power demand of supercomputers for FP64 ExaFLOPS, at least it is unlikely under reasonable power consumption, therefore, it is necessary to pair the CPU with the GPU or other accelerators. It is expected that the supercomputing system may use AMD’s Instinct MI250X or a successor model, and may also add other accelerators.
