Promised for 2027 racks, mixing Nvidia and AMD silicon in one liquid-cooled box
HPE's next-gen Cray supercomputing platform will offer a choice of compute nodes with Nvidia's Vera Rubin or AMD's upcoming Venice Epyc CPUs – or a mix of both.
The HPE Cray Supercomputing GX5000, some details of which were disclosed last month, is the next iteration of the firm's architecture for high performance computing (HPC) systems.
On Thursday, the firm unveiled details of three compute blades that will initially be offered, the server nodes that will fill the liquid-cooled compute racks of any future deployment, new management tools, and two new supercomputers for German research bodies.
None of the new kit is expected to be available until early 2027, save for the Cray Supercomputing Storage Systems K3000 announced last month, which will ship in early 2026.
Of the compute blades, the most notable is the Nvidia-based GX440n Accelerated Blade, which will feature 4 Vera CPUs and 8 Rubin GPUs for "mixed-precision computing," which we take to mean AI as well as more traditional HPC workloads.
Each Vera CPU comprises 88 custom-designed Arm cores, while Rubin is the successor to Nvidia's current Blackwell GPUs, scheduled for release in 2026.
Each Cray rack can be configured with a maximum of 24 of these blades for up to 192 Rubin GPUs per rack, HPE said.
The remaining two blades will be based on AMD's sixth-gen Epyc processors, codenamed Venice, which are also expected to launch in 2026 and understood to feature up to an impressive 256 CPU cores each.
The GX350a Accelerated Blade is similarly aimed at mixed-precision computing, with one Venice CPU and 4 of AMD's Instinct MI430X series GPUs. Up to 28 of those blades can be configured per rack, for a total of 112 MI430X accelerators.
For customers wanting to build a CPU-only cluster, the GX250 Compute Blade fits eight of those AMD Venice chips, and up to 40 of those blades will fit in each rack, enabling up to 81,920 CPU cores.
HPE said that a mix of these nodes can be fitted in a Cray GX5000 compute rack (pictured below), to enable the most cost-effective combination of processing blades for smaller deployments - the first time it has been possible to mix nodes this way in a Cray deployment.
Networking for systems based on this kit will be the 400 Gbps version of HPE's Slingshot interconnect, which can be regarded as a high-performance version of Ethernet.
All of the blades can be configured with either four or eight Slingshot ports per blade via a liquid-cooled PCIe mezzanine network card, while all the blade types can optionally fit a pair of NVMe SSDs.
Also new is a liquid-cooled Slingshot 400 switch blade featuring 64 network ports, plus a blade chassis that can fit eight, 16 or 32 of these blades.
The Cray Supercomputing GX5000 racks themselves are designed to accommodate up to 400 kilowatts of equipment, with the potential to go up to 1 megawatt in the future, HPE claims. Each holds up to 44 of the slightly taller Open Compute Project (OCP) specification rack units, which are 48 mm high to accommodate power shelves, compute blades, or Slingshot switch blades.
Two customers lined up to get supercomputers based on the new kit are the University of Stuttgart's High Performance Computing Center (HLRS), and the Leibniz Supercomputing Center (LRZ) of the Bavarian Academy of Sciences and Humanities.
However, with the chips to power these not due to launch until next year, there's not much info yet about them. We know the Stuttgart system will be called Herder and is expected in 2027, and the LRZ computer will be known as Blue Lion and is expected to deliver sustained performance up to 30 times faster than the current setup. That's about it.
HPE also announced Supercomputing Management Software to enable customers to deploy converged AI and HPC workloads on its Cray hardware.
This supports multi-tenant, virtualized, and containerized environments and also provides system-wide management of power. The latter enables monitoring of energy consumption and integrates with power-aware schedulers to maximize energy efficiency and forecast running costs, the firm said. ®