By Anton Shilov published 2 days ago
Atos’ BullSequana XH3000 to enable exascale-class supercomputers in Q4 2022
Atos has introduced its new BullSequana XH3000 hybrid computing platform that enables exascale-class supercomputers. The new high-performance computing platform is agnostic to the hardware architectures it uses, so it is compatible with CPUs from AMD, Intel, Nvidia, and SiPearl, as well as compute GPUs from AMD, Intel, and Nvidia. In addition, the platform uses liquid cooling and can handle nodes with up to 1000W power consumption. The first systems are due in Q4.
Atos’ BullSequana XH3000 hybrid computing platform will offer maximum flexibility for systems with performance from 1 FP64 PetaFLOPS to 1 FP64 ExaFLOPS (and to 10 FP8/FP16 ExaFlops for AI applications) leveraging hardware due to be released during the next six years. In addition to traditional x86 CPUs from AMD and Intel, it also supports upcoming Arm-based Grace processors from Nvidia and Rhea high-performance system-on-chips from SiPearl. Furthermore, as suggested by its hybrid nature, the platform can support next-generation compute GPUs (or accelerators), including AMD’s CDNA 2/3-based designs, Intel’s Ponte Vecchio, and Nvidia’s next-gen GPU solutions.
The future CPUs and GPUs for datacenters and high-performance computing (HPC) promise to be extremely power-hungry, which is why the BullSequana XH3000 is designed for liquid cooling by default. The XH3000 will use Atos’ Direct Liquid Cooling (DLC) solution that promises to provide ‘over 50% more cooling power than previous generations’ and enable nodes that consume up to 1000W. While 1kW of power per node may sound extreme, it looks like this will be a fairly standard energy-efficient HPC node configuration in the future. For example, rumor has it that AMD’s Instinct MI250X consumes up to 550W, whereas the industry-standard OAM form-factor for such units is designed to supply up to 700W of power.
At present, we do not know a lot of specifics about the upcoming HPC solutions from AMD, Intel, Nvidia, and SiPearl, but the very fact that Atos claims compatibility with them indicates that their general characteristics (e.g., power consumption) are known to the server maker. In fact, Nvidia states that it has worked with Atos to build the platform.
“By combining the well-known expertise Atos has with Nvidia AI and HPC technologies and work at our joint lab, this platform will allow researchers to get significant insights much faster to grand challenges both in supercomputing and industrial HPC,” said John Josephakis, global vice president of sales and business development for HPC/supercomputing at Nvidia.
Given the availability timeframe of the first BullSequana XH3000 machines, it is likely that they will be among the first ones to support Nvidia’s upcoming next-gen compute GPUs due later this year.
As for interconnections, the BullSequana XH3000 supercomputer architecture will support a wide range of networking technologies, including BXI, High Speed Ethernet as well as HDR & NDR InfiniBand.
All three U.S. supercomputers announced to date rely on HPE’s Cray EX architecture that can use almost any CPU and various compute accelerators. These days Cray EX systems use AMD’s or EPYC or Intel’s Xeon Scalable CPUs along with compute GPUs from AMD, Nvidia, or Intel (when Ponte Vecchio is available). By adding support for Nvidia’s Grace and SiPearl’s Rhea, Atos offers a more flexible platform (yet we do not know what can stop HPE from adopting its architecture for Grace or Rhea SoCs).
“We are extremely proud of our role as a leader in HPC and of our new BullSequana supercomputer, revealed today, which results from 15 years of R&D efforts and brings together Atos’ proven expertise and experience in high-performance computing, AI, quantum, security and digital decarbonization,” said Rodolphe Belmer, CEO of Atos. “It will no doubt enable, through the gateway of exascale, some of the key scientific and industrial innovation breakthroughs of the future.”