Future exascale supercomputers will be built using these HPE Lego-like cases that could house nearly 100,000 AMD EPYC cores
- HPE’s Cray EX4000 supports modular designs, scalable to 98,304 cores
- Slingshot 400 interconnect doubles throughput, essential for exascale systems
- Thanks to the collaboration between AMD and Nvidia, the computing density can exceed 10 petaflops
Future exascale supercomputers will adopt modular designs that enable improved scalability and efficiency, new reports claim.
The NextPlatform says Hewlett Packard Enterprise (HPE) is leading this transformation, with its Cray EX4000 systems, which leverage HPE’s Slingshot interconnect and innovative chassis to support up to 98,304 AMD Epyc cores per rack.
HPE’s interconnect technology is also advancing. Expected to be deployed in “Shasta” Cray EX systems in 2025, the Slingshot 400 builds on the previous Rosetta architecture and promises 400 Gb/sec per port, doubling the throughput of its predecessor.
Stay ahead
This development is critical for managing the massive data loads produced by exascale systems, especially for AI and scientific simulations.
The Slingshot 400’s improvements to network interface cards and switches demonstrate HPE’s commitment to remain competitive in a market that includes Nvidia and Broadcom, and The NextPlatform predicts that the company could “achieve speeds of 800 Gb/sec with a future Slingshot 800 (perhaps in fall 2027) and 1.6 Tb/sec speeds with an even more distant Slingshot 1600 (perhaps in fall 2029) .”
On the compute side, HPE’s partnership with AMD is reflected in the upcoming Cray EX4252 Gen 2 compute blades. These will feature AMD’s Zen 5c-based Epyc 9965 processors, which offer 192 cores per CPU.
According to The NextPlatform“With eight Epyc 9965s on each blade and 64 blades in a single EX4000 chassis, that amounts to a total of 98,304 cores in a single liquid-cooled rack.” This setup achieves a maximum of 2 petaflops per rack, and with 500 racks “you could build an exascale CPU-only supercomputer.”
Nvidia of course also plays a key role in HPE’s systems. The upcoming EX154n blade, expected in late 2025, will feature Nvidia’s Grace-Blackwell complexes, combining CPUs and GPUs in a unified design. A single cabinet using these components can hold more than 10 petaflops.
HPE’s advancements, which were presented at the SC24 supercomputing conference, include storage upgrades. The new Cray E2000 all-flash array doubles the performance of its predecessor with PCIe 5.0-compatible NVM Express drives, ensuring storage keeps pace with computing and networking developments.
By upgrading its range from top to bottom and combining the very latest computing, networking and storage technologies, HPE, already a leader in the HPC server market, is demonstrating its determination to push the boundaries of high-performance computing and expand its maintain dominance over the market. like rivals Dell and Lenovo.