Nvidia Arm Superchips to Power $160 Million Supercomputer in Barcelona

Nvidia Arm Superchips to Energy $160 Million Supercomputer in Barcelona

Posted on

Nvidia’s Grace superchip made waves when launched earlier this 12 months, as the corporate promised a supercharged Arm-based product that might tackle Intel and AMD’s x86 dominance within the Excessive-Efficiency Computing (HPC) area. Now, as reported by HPC Wire, the corporate has snagged a $160 million contract (~€151 million) to offer the brains and brawn of supercomputing {hardware} for certainly one of EuroHPC’s supercomputing tasks. The MareNostrum 5 (MareNostrum roughly interprets to “our sea”) can be put in within the Barcelona Supercomputing Centre (BSC) in Spain and can be operational as early as 2023.

Mare Nostrum 5 is being constructed as a part of the EuroHPC JU mission, and is predicted to supply peak efficiency of 314 Petaflops of FP 64 computing energy throughout each CPU and GPU accelerators, with 200 Petabytes of storage for in-access workloads, and an additional 400 Petabytes of chilly storage. Following traits in HPC structure design and different tasks throughout the EuroHPC mission, it is anticipated that the 200 Petabyte node can be saved in a quick, NAND-based storage subsystem, whereas the chilly storage node (additionally referred to as lively storage, referring to knowledge that is essential however not steadily accessed) will possible make use of less expensive, classical HDD topologies.

Nvidia’s dual-Grace superchip board design. Based mostly on TSMC’s  5nm manufacturing course of, Grace helps all the newest connectivity tech, like PCIe Gen 5.0, DDR5, HBM3, CCIX 2.0 and CXL 2.0. (Picture credit score: Nvidia)

The system will make use of Nvidia’s 144-core, Arm-based Grace “superchips” in dual-chip configurations, paired with the corporate’s H100 (Hopper) discrete GPU accelerators (which characteristic 80 billion transistors apiece with 80 GB of HBM3 reminiscence and three.2 TB/s bandwidth). Because of this, MareNostrum 5 is projected to ship greater than 18 Exaflops of AI acceleration (usually FP8 8-bit floating-point operations), making it the quickest AI supercomputer within the European Union. Moreover Nvidia’s chip tech, the corporate’s Quantum-2 (aka NDR) InfiniBand software-defined networking (opens in new tab) leveraging the corporate’s BlueField knowledge processing items (DPU) to maintain all parts speaking at low latency with a excessive throughput of 400 GB/s – not not like the efficiency achieved by Cray’s Slingshot interconnect.

Supply hyperlink

Leave a Reply

Your email address will not be published. Required fields are marked *