HPE will deliver two new supercomputers to Oak Ridge National Laboratory. Discovery is scheduled to arrive in 2028 and enter operations in 2029.

The Oak Ridge award gives HPE a second flagship deployment for Cray technology, five years after Frontier reached the lab on the Cray EX2000 platform. Discovery will run the all-flash GX5000 alongside a K3000 DAOS storage tier and the existing Lustre-based E2000. A second system, Lux, will provide a dedicated AI factory for U.S. researchers on AMD Instinct MI355X GPUs.

What does the GX5000 change for ORNL?

Discovery moves ORNL from the EX2000 compute fabric that powers Frontier to the all-flash GX5000, a more compact Cray design that needs 25 percent less data center space per rack than the EX4000 variant. The new system uses HPE's Slingshot 400 interconnect. It carries 400 Gbps per port from a 51.2 Tbps switch ASIC and replaces the Slingshot 200 fabric that Frontier runs.

Frontier now sits at number two on the June 2025 TOP500 list, after Lawrence Livermore's El Capitan took the top position. ORNL director of science Bronson Messer tied the two new systems to a broader productivity gain. "We expect both systems will contribute to a paradigm shift in our productivity, reaching unparalleled gains in various, critical areas of scientific research and leadership."

Takeaway: Discovery cuts rack footprint by a quarter and doubles the interconnect line rate, while keeping Frontier's storage tier in place.

How much faster is the storage?

A fully populated GX5000 rack delivers up to 75 million IOPS, against 54 million for a Cray E2000 all-flash rack and 18 million for Frontier's E1000 subsystem. That puts the new platform 39 percent ahead of the E2000 and more than four times above the storage tier ORNL runs today. The K3000 turns DAOS into a product rather than an integration project, the first factory-built configuration based on Intel's Distributed Asynchronous Object Storage.

Takeaway: Discovery pairs a 39 percent IOPS jump over the E2000 with a four-times lift over Frontier's existing storage.

Why does a factory-built DAOS tier matter?

HPE will validate the hardware, firmware, and storage node design before shipment, replacing the on-site assembly that DAOS deployments have required to date. DAOS-backed systems already hold the top two positions on the IO500 storage benchmark — Aurora at Argonne National Laboratory and SuperMUC at Leibniz Supercomputing Center — with a combined score four times higher than the next thirty systems.

A K3000 storage node is a 1 RU HPE ProLiant DL360 Gen12 server with twenty EDSFF bays for NVMe PCIe Gen 4 E3.S SSDs. A rack holds between four and forty nodes, with an average configuration of twenty. Forty nodes deliver a maximum raw capacity of 12.32 PB at 15.4 TB per drive, and three SSD capacities are offered — 3.84 TB, 7.68 TB, and 15.4 TB. HPE senior distinguished technologist Johan Lombardi will discuss the DAOS 2.8 release at the DAOS User Group meeting on November 16, adjacent to Supercomputing 2025.

Takeaway: DAOS becomes a shipped product with validated hardware, replacing the on-site assembly that previous deployments required.

What does the AI half of the contract do?

The Lux system gives ORNL a separate, multi-tenant AI factory for training and inference. It runs AMD MI355X accelerators on the ProLiant XD685 platform with EPYC host CPUs and Pensando fabric. HPE president and CEO Antonio Neri framed the deal as a continuation of the Frontier-era partnership. "When we built Frontier for Oak Ridge National Laboratory and ushered in exascale, we achieved the pinnacle in supercomputing history and a triumph for the U.S. We are proud to build on that leadership innovation and strong public-private partnership with the U.S. Department of Energy, ORNL and AMD, to build Discovery and Lux, accelerating the next era of scientific discovery and AI innovation."

Discovery will pair the GX5000 with the new K3000 DAOS tier and the Lustre-based E2000. That gives ORNL a tiered storage stack rather than a wholesale replacement. Takeaway: the AI factory arrives as a separate machine rather than a partition of the HPC system, leaving Lustre and DAOS to share the storage tier.