Arm’s journey into the HPC market reached several significant milestones this week at SC18. While Arm has worked diligently on our HPC software stack, it is hugely gratifying to see the quality of those efforts. A common theme among all the deployment announcements was “it just works”. While we expect further great things from Arm and our partners in the years to come, let’s take a moment to highlight some of the key news and sightings from SC18 in Dallas, TX.
Kicking off the SC week was a significant milestone for the Arm HPC business, as Sandia National Labs announced Arm’s first entry into the Top500 List for supercomputers. This is an HPC industry-wide list that tracks the performance (based on the HPL benchmark) of all the top supercomputers in the world. The Sandia Astra supercomputer, based on the HPE Apollo 70 and the Marvell ThunderX2 SoC, was just recently deployed and showed up at a respectable #203 out of 500.
In Europe Arm’s HPC business also took a step forward. Cray announced that full deployment had occurred for the UK’s GW4 Isambard supercomputer. This system is based on the Marvell ThunderX2 version of the Cray XC50. It is currently the largest Arm-based supercomputer in Europe, though we are not certain how long that title will last as CEA also announced plans to deploy a sizable ATOS BullSequana supercomputer by early 2019.
We also saw strong interest continue from major US labs as Los Alamos National Labs announced collaboration with Marvell and Cray and the live deployment of the “Thunder” supercomputer based on the Cray XC50. Early feedback from our DOE partners highlights a rapidly maturing Arm HPC software ecosystem and a strong compute efficiency (as a percentage of peak).
While a cold front had chilled Texas the entire week of SC18, the HPC Cloud space was heating up for Arm. Announcements such as Verne Global adding Marvell ThunderX2 to their hpcDIRECT bare-metal cloud service, and HPC Systems launching their Arm-based Science Cloud in Japan, hosting both Marvell and Huawei servers based on the Arm architecture. Having cloud accessible resources for ISV’s, ecosystem developers, and end users should help accelerate Arm deployments further for HPC.
Fujitsu provided an update on their POST-K efforts at the SC18 Exhibitor’s Forum. They divulged further information on their performance and compute density, noting 1PFLOP of double-precision compute per rack. Furthermore, Fujitsu POST-K is on schedule for its production deployment in early 2021.
For the first time at SC, Huawei was showing off Arm-based servers alongside the x86 boxes in their booth on the exhibition floor. The Huawei TaiShan V1 and TaiShan V2 servers were on-hand for the public in familiar two-socket configurations and showed promising workload performance with their strong memory capacity and throughput.
Arm hosted its fourth annual HPC User Group event this past week. The event hosted a collection of presentations and Q&A sessions from key end-user sites world-wide. For the first time at SC, our agenda was filled entirely by end users, leaving no additional room on the agenda for technology partners. Content covered end users’ experiences, plans, performance tuning and comparisons, as well as status of deployments happening across the globe. We had ~175 folks turn up for the sessions, but if you were one of those that couldn’t make it there, we will have the PDF’s posted onto our HPC developer site shortly.
Please see our dedicated HPC website for further HPC ecosystem news and information.
Arm HPC website