Europe’s sovereign approach to exascale computing is complicating plans for U.S. chipmakers to break through in the market — and, in the process, empowering local chipmakers.
For one, European chip startup SiPearl is emerging as an early beneficiary amid efforts by the U.S. and EU to weaponize semiconductors and create the world’s fastest computers.
SiPearl, which is based in France, is becoming a go-to company for the world’s top chipmakers to pack proprietary accelerators in Europe’s upcoming top-end systems. The France-based company is at the spearhead of Europe’s plans to develop primary processors for exascale, with SiPearl’s made-in-Europe Rhea CPUs on EU’s roadmap for future exascale computers.
Like the U.S. and China, Europe is seeking chip independence as governments turn semiconductors into political bargaining chips. The EU is intensifying efforts to make supercomputers and exascale systems with homegrown processors and components, while cutting reliance on foreign technology.
Two of the fastest European supercomputers on the Top500 list – Lumi and Leonardo – are based on proprietary x86 chips from U.S. companies AMD and Intel. With a shift to Rhea, Intel, AMD and Nvidia see SiPearl’s chip as a gateway to put their GPUs and other accelerators in EU’s exascale systems.
SiPearl already has partnerships with Intel and Nvidia, and this week announced a partnership with AMD, which wants to expand the market for its Instinct GPUs to European supercomputers.
AMD’s Instinct GPUs already powers the world’s first exascale system, Frontier, at Oak Ridge National Laboratory. The SiPearl-AMD partnership revolves on making Rhea compatible with AMD’s Instinct accelerators by improving the ROCm parallel programming framework. AMD’s enterprise GPUs are compatible with x86 chips, and the focus will be on adding Arm-based compatibility with the GPUs.
“You have to spend lots of time in terms of integration testing and optimization. That’s what we do now with OneAPI [with Intel] and Nvidia with CUDA,” said Philippe Notton, CEO of SiPearl, in an interview with HPCwire at SC22 in Dallas. He added that since Nvidia has done some of the GPU porting on Arm already, it should make Nvidia’s GPU compatibility with Arm Rhea much easier.
The partnership with AMD will help SiPearl provide a wider range of GPUs along with Rhea to its high-performance computing customers, Notton said, adding that the AMD partnership is very similar to what SiPearl did with Intel on OneAPI.
“We have a dedicated team on both ends to ensure that basically our chip and Intel’s chip can work [with] OneAPI. And that is what we have just announced for AMD,” Notton said.
Beyond hardware, the road for the EU to exascale includes tools, compilers, runtimes and system integration tools for chips and accelerators. The EU is funding multiple efforts that include the EPI (European Processor Initiative), EUPEX (European Pilot for Exascale) and EuroHPC. Participants in these efforts include academics, researchers and European commercial organizations such as Atos, which plans to build exascale systems under its BullSequana line.
The Rhea processor is at the center of EU’s blueprints to create exascale systems, and software compatibility with Rhea is important for Nvidia, AMD and Intel to win more European exascale business. The first exascale system based on the Rhea CPU could go live in 2023 or 2024, according to a roadmap published by EPI.
EUPEX has built reference systems around Rhea processors, Atos’ BXI (BullSequana eXascale Interconnect) switches and OpenSequana racks. Each rack has up to 96 Rhea GPUs and 32 GPUs. There is no clarity on what GPUs would be used in the EUPEX systems.
The first European exascale supercomputer is expected to be JUPITER (Joint Undertaking Pioneer for Innovative and Transformative Exascale Research), which will be installed at the campus of Forschungszentrum Jülich (FZJ) and will be operated by the Jülich Supercomputing Centre. The system is expected to go live in 2023, and Jülich is sending out RFPs as part of its hardware procurement program.
JUPITER’s hardware specifications are not yet clear, and it is not certain if the system will be built using sovereign European technology. The system is being built ahead of the EU’s plans to develop an all-Euro supercomputer.
There is a desire to shift to made-in-Europe technologies in JUPITER, but its final configuration depends on the kind of proposals provided by the hardware makers, Estela Suarez, head of RG next-generation architectures and prototypes at Forschungszentrum Jülich, told HPCwire on the conference floor.
“This system is half financed by EuroHPC, and that’s definitely on the agenda to try to have as much European technology as possible on the system. At the end, it also depends on what is offered within the procurement, what the vendor brings… and how things turn out. But yes, we would definitely like to have some European technology features,” Suarez said.
The Jülich Supercomputing Center is already testing a range of chips that include chips from Intel and Nvidia, and computers that include quantum annealing systems from D-Wave.
SiPearl was founded in 2019 with funding from the European Union. Today, the company has offices in six locations, and though it started with EU funding, it is a for-profit organization.
“The way you manage a chip company is you need to go quite high in terms of value. If you just sell a chip without the software, you’re dead. The more you do, the better it is,” Notton said.
The EPI’s general-purpose processor roadmap includes a Rhea2 being made using the 5nm process in 2024, and a third-generation chip beyond 2025. The Rhea chip is based on Arm – which was previously European tech – and is also being viewed as a stopgap technology until European companies are realistically able to switch over to RISC-V, which is an instructional set architecture that is free to license. The Rhea chips mix Arm CPUs with RISC-V controllers.
The RISC-V architecture is not yet ready for high-performance computing due to software, compatibility and other issues, said Krste Asanović, a professor at the University of California, Berkeley and one of developers of RISC-V, during a speech at the Supercomputing 2022 conference.
SiPearl’s Notton said that licensing an Arm design was the quickest way to stand up an architecture to help Europe meet its goals to develop a sovereign chip.
“Arm is the only core that is competitive, beyond x86, if you want to have a quick time to market to be able to develop a chip in three years and not in 10 years,” Notton said.
While the EU has its own CPU, it may need to rely on U.S. companies for accelerators to reach its exascale goals. The EPI has a parallel effort underway to develop high-performance accelerators, though it may be years until it develops a competitive product.
EPI’s high-performance accelerator, called EPAC, is based on RISC-V, which meets the European goal of being non-proprietary and easy to replicate. EPI also wants to make sure that its homegrown accelerators can easily connect to other chips developed by EPI.
“We should not transmit the message that EPI is getting the next Nvidia or whatever. This requires a lot of time, a lot of technology. Maybe we will arrive there or not,” Filippo Mantovani, senior researcher at Barcelona Supercomputing Center, told HPCwire on the show floor.
The bigger goal is to create a strong knowledge base in creating chips and accelerators, which is important in developing European expertise and a thriving ecosystem in the region, said Mantovani, who is also a leader in EPI’s effort to develop accelerators.
“What is really important is that we have European support for building the whole chain of knowledge that you need to make good chips. It is not enough to have somebody with a good architectural idea on a blackboard unit. You need the whole chain of knowledge – from the architecture, to the mapping, to the technology, to the tape out, to the compiler, to the software and so on,” Mantovani said.