The future exascale supercomputers will be built with HPE modular cabinets that can accommodate nearly 100,000 AMD EPYC cores.
HPE has developed the only three exascale supercomputers in the world so far.
Future exascale supercomputers will adopt a modular design that will allow for improved scalability and efficiency, according to recent reports. Hewlett Packard Enterprise (HPE) stands out in this transformation with its Cray EX4000 systems, which integrate HPE's Slingshot interconnect and an innovative chassis that supports up to 98,304 AMD Epyc cores per rack.
HPE's interconnect technology is also advancing significantly. The Slingshot 400, set to be used in the Cray EX "Shasta" systems in 2025, is expected to surpass the previous Rosetta architecture by delivering performance of 400 Gb/sec per port, effectively doubling the performance of its predecessor. This advancement is crucial for handling the massive data loads generated by exascale systems, particularly in artificial intelligence applications and scientific simulations.
Improvements in the Slingshot 400's network interface cards and switches reflect HPE's commitment to staying competitive in a market that includes Nvidia and Broadcom. It is speculated that HPE could achieve speeds of “800 Gb/sec” with a future Slingshot 800, possibly in the fall of 2027, and “1.6 Tb/sec” with the Slingshot 1600, anticipated for the fall of 2029.
In terms of computing, HPE's collaboration with AMD is materializing in the upcoming Cray EX4252 Gen 2 blades. These will incorporate Epyc 9965 processors, based on the Zen 5c architecture, offering 192 cores per CPU. According to reports, with eight Epyc 9965 processors in each blade and 64 blades in an EX4000 chassis, a total of 98,304 cores can be achieved in a single liquid-cooled rack, reaching up to 2 petaflops per rack, which would enable the construction of a complete CPU-based exascale supercomputer with 500 racks.
Nvidia will also play a key role in HPE's systems. The EX154n blade, expected by late 2025, will include Nvidia's Grace-Blackwell complexes, which combine CPUs and GPUs in a unified design. It is estimated that a single cabinet using these components could exceed 10 petaflops.
HPE, already a leader in the HPC server market, has also shown advancements in storage that will be presented at the SC24 supercomputing conference. New Cray E2000 all-flash arrays double the performance of their predecessor thanks to PCIe 5.0-enabled NVM-Express drives, ensuring that storage keeps pace with advances in computing and networking. By revamping its product line from top to bottom and integrating cutting-edge computing, networking, and storage technologies, HPE demonstrates its determination to push the boundaries of high-performance computing and maintain its dominance against competitors like Dell and Lenovo.
