Deal involves AMD’s Instinct GPU accelerators, which are also used in Frontier, the fastest supercomputer in the world.
AMD has announced plans to work with French chip designer SiPearl to build exascale-level supercomputing systems that use SiPearl’s Arm-based Rhea processor with AMD’s Instinct GPU accelerators.
SiPearl is a relatively small startup that began operation in 2019 with a license for Arm’s Neoverse high-performance technology. It has forged a number of alliances with partners including Intel, Nvidia, HPE, and Graphcore.
SiPearl is also involved with the European Processor Initiative (EPI), a consortium selected by the European Union to support the development of a European microprocessor specifically for high performance computing (HPC), as well as emerging applications such as artificial intelligence. The EPI’s goal is to develop an Arm-based processor for an exascale supercomputer by 2023.
SiPearl’s Rhea microprocessor was designed in partnership with French IT giant Atos. It comes with 72 next-generation Arm Neoverse cores and supports both DDR5 and high-bandwidth HBM2E memory interfaces. This gives it both extremely fast memory (the HBM) and very large memory pools (the DDR5). The processor will be manufactured by TSMC and is expected ship in 2023.
The Instinct accelerators are based on AMD’s Radeon GPU technology. They can be found in, among other systems, Frontier, an exascale supercomputer at the Oak Ridge National Labs that is the fastest supercomputer in the world.
AMD and SiPearl said they will initially start working on the interoperability of AMD’s ROCm software stack with SiPearl’s Rhea microprocessor. ROCm is AMD’s software stack for developing GPU-accelerated applications. This is a form of future-proofing the software ecosystem so future apps will run on supercomputers to be developed by the two firms.
This joint work will start by focusing on porting and optimization of AMD’s Heterogeneous Interface for Portability (HIP), AMD’s GPU programming environment that lets developers write code that can be run on both AMD and Nvidia GPUs.
“Providing a broader choice for European supercomputer end-users, it will enable Europe to tackle the great challenges of our time such as artificial intelligence, climate modelling and medical research,” said SiPearl CEO and founder Philippe Notton in a statement.