The ISC conference is here once again. This annual meeting reminds us of the importance of HPC to us all, and to Europe in particular. HPC powers the industrial heart of Europe, it contributes to almost every scientific endeavor in its world leading research institutions. In our current crisis – it is helping to keep us safe through sequencing and tracking of virus mutations, advancing vaccine development and epidemiological modeling.
Arm’s role in HPC has made some significant leaps in the last twelve months. It is exciting to see the increased impact of our architecture and our partners on this scientific and industrial community - and the developments keep coming.
SiPearl, the technology company that aims to rebuild Europe’s position and capabilities by developing its own processor, is progressing well. Within Europe, we are collaborating with many European research institutions to prepare applications, tools, and toolchains for Exascale and to look beyond.
The RIKEN Fugaku system entered official production use, with over 7.5 million Fujitsu A64FX Arm architecture cores. Having won the coveted top 500, twice, it has already produced impactful work on proteins, aerosol dispersion and virus structure for COVID-19. This will power research in Japan for many years to come. Portugal’s Minho Advanced Computing Centre will commission the EuroHPC JU Arm system in the coming months, based on Fujitsu A64FX.
This spring saw the announcement of another future Arm processor, NVIDIA’s Grace CPU, and its selection by one of Europe’s leading HPC centers, CSCS, to power their next leadership class system.
Beyond flagship HPC centers – the design choices made by our partners have brought highly competitive HPC application runtime and are often leading significantly from a cost per simulation perspective. HPC use in the cloud continues to grow. Oracle Cloud Infrastructure recently added Ampere CPUs to their platform. The Ampere Altra SoC offers 80 or 128 Arm Neoverse N1 cores per socket. AWS’s Graviton 2 Arm architecture processors already host a significant number of HPC users – and have the fastest interconnect across the AWS fleet – and up to 40% cost advantages over other processors.
We announced our next generation of processor IP in April: The Neoverse V1 and N2 cores. These cores are optimized for highest performance and highest throughput respectively. Furthermore, we launched the CMN-700, the next generation of our coherent mesh network which supports up to 256 cores per die. Relative to the Neoverse N1, the Neoverse V1 with 2x256 bit scalable vector extension (SVE) units, increases vectorization potential and capacity which translates to some significant performance gains. HACCmk, for example, sees a 6.9x performance uplift. New int8 and bfloat16 SVE instructions provide a 4x uplift per cycle for some key ML operations.
This IP is at the center of some significant developments across the globe. The development of technology, capability, and skills closer to home is the intent of many nations and regions as they seek to build better economic prospects. Arm is at the heart of enabling this through licensing our IP and our architecture (ISA).
The Neoverse V1 is the core that powers Rhea – the first SoC for HPC from SiPearl.
In South Korea, ETRI is building the K-AB21 SoC with the Neoverse V1 and their own XPU many-thread AI/HPC cores. More recently C-DAC and MEITY announced their plans to build India’s first Exascale computing platform using Arm Neoverse V1 – which is enabling India to execute on a major government priority.
Once more a virtual event, I shall miss the special atmosphere in Germany this year – particularly that the bi-annual alignment of stars when international HPC and international football coincide. Two exciting championships – the Euros and the top 500 – and we wish the participants of each the best of success.