Hewlett Packard Enterprise (HPE) announced a broad expansion of its NVIDIA AI Computing by HPE portfolio. The updates introduce new hardware based on the NVIDIA Vera Rubin architecture and specialised security features for sovereign and regulated environments.
HPE is among the first to offer compute blades featuring the NVIDIA Vera CPU. The new HPE Cray Supercomputing GX240 compute blade supports up to 16 Vera CPUs per blade. This system uses 100% fanless direct liquid cooling to manage heat while maintaining high hardware density.
A single rack of the GX240 can scale to 640 Vera CPUs and 56,320 ARM-compatible cores. This infrastructure is built for the most intensive tasks, such as training trillion-parameter large language models (LLMs) and running agentic AI, autonomous agents that perform complex reasoning and task execution.
Scaling the AI factory
The HPE AI Factory portfolio now includes the NVIDIA Vera Rubin NVL72 rack-scale system. This configuration integrates 36 Vera CPUs and 72 Rubin GPUs. Compared to previous Blackwell-based systems, this architecture offers up to a 10X reduction in costs for inference tasks and requires four times fewer GPUs to train Mixture-of-Experts (MoE) models.
HPE also introduced the HPE Compute XD700, an AI server based on the NVIDIA HGX Rubin NVL8. This system provides double the GPU density of previous generations, supporting up to 128 GPUs per rack.
Advancements in Supercomputing
For research institutions and sovereign entities, new blades have been introduced for the Cray Supercomputing GX5000 platform. These include:
- Vera CPU Compute Blade: The GX240 blade supports up to 16 Vera CPUs. A single rack can scale to 640 CPUs and 56,320 Arm-compatible cores.
- Rubin NVL72 System: This rack-scale system supports 72 Rubin GPUs. It is designed for models with over one trillion parameters, offering a 10X reduction in costs for inference tasks compared to previous systems.
- Liquid Cooling: The infrastructure uses 100% fanless direct liquid cooling. This design reduces power consumption by 65% and allows for higher hardware density in data centers.
Security for sovereign and private clouds
In terms of meeting the needs of regulated industries like defence and finance, HPE Private Cloud AI now offers an air-gapped configuration. This ensures that sensitive data remains isolated from external internet connections.
The hardware integrates specialised data processing units (DPUs) and high-speed Ethernet networking. These components offload tasks from the main processors, which improves overall system speed. On the security front, the systems include post-quantum cryptography and FIPS 140-3 Level 3 certification to protect sensitive data against emerging digital threats.
The private cloud system can now scale up to 128 GPUs through new network expansion racks. It also integrates NVIDIA Mission Control software, which automates workload orchestration and system recovery. Security is further bolstered through a partnership with CrowdStrike and support for Fortanix Confidential AI, protecting data during processing.