HPE has introduced a portfolio of advanced supercomputing servers powered by GPUs from industry leaders Nvidia and AMD. Central to this offering is HPE's recently revealed fanless direct liquid cooling system architecture, boasting a significant 37 per cent reduction in cooling power per server blade compared to traditional hybrid direct liquid cooling.
At the top of the range are state-of-the-art Cray EX solutions: the EX4252 and EX154n. The EX4252, a one-rack unit equipped with eight of AMD’s cutting-edge 5th gen EPYC processors, delivers an astounding 98,304 cores in a single cabinet.
The EX154n, designed to accommodate up to 224 Nvidia Blackwell GPUs, features the formidable Nvidia GB200 Grace Blackwell Superchip, dramatically shortening supercomputing workload completion times.
The EX4252 is slated for a Spring 2025 release, whereas the EX154n will debut towards the end of 2025, aligning with Nvidia's production timeline for Blackwell GPUs.
In addition to these revolutionary servers, HPE has introduced the Slingshot interconnect 400, promising double the line speed of its predecessor and enhanced capabilities such as automated congestion management and adaptive routing for ultra-low tail latency. This upgrade enables users to manage large-scale workloads with markedly reduced network infrastructure.
Further expanding their Cray Supercomputing portfolio, HPE unveiled the Storage Systems E2000, a high-performance storage solution for large-scale supercomputers. The E2000 reportedly doubles I/O performance compared to existing solutions, enhancing CPU and GPU utilisation by minimising idle time during I/O operations. The E2000 will be available in early 2025, with the Slingshot Interconnect 400 following later in the year.
HPE also launched the Cray Supercomputing User Services Software, which aims to optimise system efficiency, regulate power consumption, and manage diverse workloads on its supercomputing infrastructure.
Complementing these developments are two new servers tailored for AI workloads: the ProLiant Compute XD680 and XD685. The XD680, capable of housing eight Intel Gaudi 3 AI accelerators in a single node, is set for a December 2024 release. The XD685, built for accelerating complex AI model training with either Nvidia H200 or Blackwell GPUs, will debut in early 2025, with an AMD Instinct MI325X variant expected in the first quarter of 2025.
HPE HPC and AI infrastructure solutions general manager Trish Damkroger said: "Our customers turn to us to fast-track their AI system deployment to realise value faster and more efficiently by leveraging our world-leading HPC solutions and decades of experience in delivering, deploying and servicing fully integrated systems."
