Hewlett Packard Enterprise is expanding its HPE Cray supercomputing lineup with new blades, storage, interconnect and management software designed to handle the rising computational and energy demands of large-scale AI and traditional high-performance computing (HPC).
HPE is positioning the updated platform as a unified architecture for research labs, sovereign computing initiatives and large enterprises that increasingly want AI and simulation workloads to coexist on the same infrastructure rather than in separate silos.
The latest additions build on last month’s introduction of the HPE Cray Supercomputing GX5000 platform and the K3000 storage system. Together, the hardware and software are meant to deliver higher compute density, better energy efficiency and more operational control as AI models grow in size and complexity. HPE argues that organizations are no longer looking just for peak performance on benchmark workloads; they are also under pressure to manage power consumption, integrate AI with existing simulation workflows, and keep infrastructure secure in multi-tenant environments.
European research centers are among the first to commit to the updated platform. The High-Performance Computing Center Stuttgart (HLRS) and the Leibniz Supercomputing Centre (LRZ) have both selected the HPE Cray GX5000 as the basis for their next flagship systems, named Herder and Blue Lion, respectively.
HLRS expects a significant jump in performance for both simulation and AI workloads while also reducing the energy footprint of its data center.
LRZ is emphasizing sustainability as much as raw performance: its Blue Lion system will use 100 percent direct liquid cooling and is being designed to run with cooling water temperatures up to 40°C, allowing waste heat to be reused across the Garching research campus. According to LRZ, the new system is projected to deliver sustained performance up to 30 times faster than its current supercomputer, while enabling tighter integration of modeling, simulation and AI.
Directly Liquid-Cooled Compute Blades
At the heart of the portfolio expansion are three new directly liquid-cooled compute blades that support different combinations of CPUs and GPUs from multiple vendors. Each blade type is intended to address a particular set of workloads, while still fitting into the same chassis and management framework so customers can mix and match based on their needs.
The HPE Cray Supercomputing GX440n Accelerated Blade targets organizations standardizing on NVIDIA platforms for mixed-precision AI and HPC. Each blade combines four NVIDIA Vera CPUs with eight NVIDIA Rubin GPUs and exposes either four or eight 400 Gbps HPE Slingshot endpoints, plus optional local NVMe solid-state storage. Up to 24 of these blades can be installed in a single GX5000 compute rack, yielding as many as 192 Rubin GPUs per rack for dense accelerator configurations.
For customers that favor AMD’s ecosystem, HPE is introducing the GX350a Accelerated Blade. It pairs a next-generation AMD EPYC processor, codenamed ‘Venice,’ with four AMD Instinct MI430X GPUs from AMD’s MI400 series. This blade is positioned as a ‘universal’ engine for AI and HPC, particularly for organizations focused on sovereign AI strategies that emphasize data locality and control. A rack can host up to 28 GX350a blades, providing up to 112 MI430X GPUs per rack.
The GX250 Compute Blade addresses CPU-only workloads that still demand high double-precision performance, such as large-scale simulations and traditional numerical modeling. Each blade carries eight next-generation AMD EPYC ‘Venice’ CPUs, delivering very high x86 core density in a single rack. In large systems, customers can combine a CPU-only partition built from GX250 blades with one or more GPU-accelerated partitions based on either the NVIDIA- or AMD-based blades, depending on their application mix and vendor strategy. Up to 40 GX250 blades can fit in a compute rack, maximizing core count per footprint.
All three blades rely on 100 percent direct liquid cooling, a design choice that reflects the broader trend toward liquid-cooled data centers as power densities climb. By removing heat more efficiently at the component level, direct liquid cooling can reduce the need for traditional air-cooling infrastructure, improve rack density and enable higher sustained performance at a given power envelope.
To manage these increasingly complex systems, HPE is rolling out new Supercomputing Management Software alongside the hardware. The platform is built to support multi-tenant environments, virtualization and containerization, allowing operators to host multiple user communities and workload types on the same infrastructure while enforcing isolation where required. Management functions span the entire lifecycle of the system, from initial provisioning to day-to-day monitoring, power and cooling control, and capacity expansions.
Interconnect Performance, Energy Awareness
A key focus of the software is energy awareness. Operators can monitor power consumption across the system, estimate usage over time and integrate with power-aware schedulers. That capability is important as both public and private operators face stricter energy budgets and sustainability mandates. HPE is also adding enhanced security controls and governance reporting to align with the requirements of sovereign computing projects and regulated industries.
Interconnect performance remains a critical factor in supercomputing architectures, especially as AI workloads become more communication-intensive. HPE is bringing its Slingshot 400 interconnect to GX5000-based systems in early 2027. Slingshot 400 is optimized for dense, liquid-cooled form factors and large, converged AI/HPC installations. The latest switch blade delivers 64 ports at 400 Gbps each and can be deployed in multiple configurations: eight switches for 512 ports, 16 switches for 1,024 ports or 32 switches for 2,048 ports. HPE says the topology is designed to exploit all available bandwidth in the GX5000 architecture and reduce latency while maintaining cost control. Slingshot 400 was initially announced for earlier Cray systems; this rollout adapts it to the denser and more AI-centric GX5000 platform.
Storage is another pillar of the updated portfolio. The HPE Cray Supercomputing Storage Systems K3000 is based on HPE ProLiant DL360 Gen12 servers and integrates the open source Distributed Asynchronous Object Storage (DAOS) stack directly from the factory. DAOS is designed for low latency and high throughput, particularly for workloads where input/output performance is as important as compute power, such as AI training pipelines and data-intensive simulations.
HPE is offering multiple DAOS server configurations, optimized either for performance or capacity. Performance-focused systems can be configured with 8, 12 or 16 NVMe drives, while capacity-optimized versions scale to 20 drives. Drive sizes range from 3.84 TB to 15.36 TB, and DRAM configurations span from 512 GB up to 2 TB per node depending on the chosen profile. Connectivity options include HPE Slingshot 200 or 400, InfiniBand NDR and 400 Gbps Ethernet, enabling integration into a variety of fabric designs.
Portfolio of HPE Services around the Cray Line
HPE emphasizes that hardware is only part of the supercomputing value proposition. The company continues to offer a portfolio of services around the Cray line, including application performance tuning, turnkey deployment and 24×7 operational support. For customers that may not have deep in-house HPC staff, these services are pitched as a way to shorten time-to-science or time-to-insight and maintain sustained performance as software stacks and workloads evolve.
The company’s partners are leveraging the announcement to underline broader industry trends. AMD highlights joint work with HPE at the convergence of HPC and sovereign AI, arguing that tightly integrated EPYC CPUs and Instinct GPUs can deliver scalable, energy-efficient systems for both scientific and AI workloads. NVIDIA stresses that the combination of its Vera Rubin platform with HPE’s next-generation supercomputers is aimed at accelerating simulation, analytics and AI in what it describes as the “AI industrial revolution.” Market research firm Hyperion Research frames the GX5000 line as part of a larger wave in which high-performance computing and AI are among the fastest-growing segments of the IT market, with direct implications for product design, scientific research and broader societal challenges.
Availability is staggered across the portfolio. The new GX440n, GX350a and GX250 blades, along with the Supercomputing Management Software and Slingshot 400 for GX5000 clusters, are planned for early 2027. The K3000 DAOS-based storage systems with ProLiant compute nodes are scheduled to arrive earlier, in 2026. That roadmap suggests that HPE is aligning its supercomputing platform with the expected next cycle of large AI and exascale-class procurements, giving early adopters time to plan architectures that blend simulation, data analytics and AI training at scale.
For B2B technology buyers and architects, the updated HPE Cray portfolio illustrates how supercomputing design is evolving in response to AI: denser blades, more aggressive liquid cooling, more flexible management of multi-tenant and containerized environments, and storage systems optimized for extreme I/O. Rather than treating AI clusters as separate, special-purpose infrastructure, the direction is toward converged systems that can run AI, simulation and data workflows side by side – sharing the same racks, fabrics and operational model.
Executive Insights FAQ: Supercomputing and Blade Architectures
Why are blade architectures so prevalent in modern supercomputers?
Blade designs make it easier to pack high core and accelerator counts into a constrained footprint while standardizing power, cooling and networking. This modularity simplifies scaling: operators can add or replace blades without redesigning the entire system, which is critical as AI and simulation workloads grow and hardware generations change more quickly.
How do blades help balance AI and traditional HPC workloads?
Blades can be specialized for different roles – GPU-heavy for AI training, CPU-dense for double-precision simulation, or I/O-optimized for data tasks – yet still operate within one chassis and management domain. That allows IT teams to build heterogeneous partitions tailored to each workload type, while presenting a unified system to schedulers and users.
What is the impact of blade density on power and cooling strategy?
Higher blade density drives up rack-level power consumption and thermal output, which is why many next-generation systems pair dense blades with direct liquid cooling. This approach can remove heat more efficiently than air, enabling sustained performance and higher rack utilization without breaching power or temperature limits.
How does using blades affect lifecycle management and upgrades?
Because compute, acceleration and sometimes even storage are modularized at the blade level, operators can phase in new CPU or GPU generations incrementally. This reduces downtime and extends the useful life of the overall system chassis, interconnect and facility infrastructure while still allowing performance upgrades over time.
Are blade-based supercomputers compatible with emerging disaggregated architectures?
Yes, blades can serve as the building blocks in more disaggregated designs where memory, accelerators or storage are pooled and accessed over high-speed fabrics. As network technologies improve, blade enclosures can evolve from tightly coupled nodes into flexible endpoints in larger resource pools, giving operators more freedom to compose systems dynamically based on workload needs.
