RISC-V chips for a digitally autonomous EU: 240 million euros for DARE SGA1

The EU project Digital Autonomy with RISC-V in Europe, Special Grant Agreement 1 (DARE SGA1), led by the Barcelona Supercomputing Center, is funded with 240 million euros. The 38 partners are developing a European infrastructure of hardware and software for high-performance computing (HPC) and AI.

Among other things, three European chiplets with RISC-V technology are planned:

• a vector accelerator (VEC) led by the Catalan company Openchip
• an AI processor (AIPU) for inferencing from Axelera AI in the Netherlands
• an HPC-optimized general purpose processor (GPP) from the Czech company Codasip

DARE-SGA1 is part of the joint EU project for high-performance computing (European High Performance Computing Joint Undertaking EuroHPC JU), which has also been funding the European Processor Initiative (EPI) and the SiPearl Rhea1, which has been postponed several times since 2018.

Samples of the Rhea1, which was originally planned for 2021, are finally due to arrive this year; the ARM chip with SVE computing units and super-fast high bandwidth memory (HBM) is to power part of the JUPITER exascale supercomputer in Jülich.

Chiplets & software for Europe

It has been clear for some time that DARE will rely on chiplets. This makes it relatively easy and flexible to combine the most suitable computing units depending on the area of application.

Existing RISC-V processor cores have so far lagged far behind the computing power of ARM and x86 designs. However, upcoming cores with the RVA23 vector extension, improved out-of-order units and higher clock frequencies could catch up. The XuanTie C930 from the chip division of Chinese cloud giant Alibaba, for example, offers RVA23 with 256 bits. And the Minerva from French company Cortus is expected to reach 4 GHz even with 22-nanometer production.

The project partners also emphasize the great importance of optimized software. Some formulations on the project page allude to the current situation in AI data centers, which the US manufacturer Nvidia dominates, thanks in part to CUDA. Although other manufacturers produce AI accelerators with theoretically higher computing power, they are not really competitive due to poorer software.

Europe is currently completely dependent on the USA for AI computing power.

(ciw)