HPE Cray Supercomputing EX4000 (NVIDIA GB200)

The HPE Cray Supercomputing EX4000 with NVIDIA GB200 is a purpose-built, liquid-cooled blade supercomputing platform engineered to maximize AI training throughput and HPC simulation density within a single rack-scale system. At its core, the EX4000 integrates the NVIDIA GB200 NVL72 architecture, pairing 36 Grace Arm-based CPU dies with 72 Blackwell B200 GPUs across a high-bandwidth NVLink Switch fabric delivering up to 130 TB/s of NVLink bandwidth within the rack domain. Direct Liquid Cooling (DLC) throughout the blade chassis enables sustained Tensor Core utilization at scale, eliminating the thermal headroom constraints that limit air-cooled GPU systems and allowing rack-level TDP ratings that far exceed traditional server designs.

The EX4000 chassis employs HPE Cray's modular blade architecture, where compute blades slot into a shared high-bandwidth backplane interconnected by a purpose-designed HPE Slingshot 11 high-speed fabric providing 200 Gb/s per port with adaptive routing and congestion management. This eliminates fabric bottlenecks that emerge in loosely coupled GPU clusters, enabling the system to operate as a unified accelerated compute fabric rather than a collection of discrete servers. The platform supports HPE Cray ClusterStor storage integration, Cray Programming Environment (CPE), and is compatible with standard MPI, NCCL, and CUDA-based HPC/AI software stacks, providing a familiar developer experience with supercomputer-grade performance characteristics.

Targeted at national laboratories, hyperscale AI research facilities, aerospace and defense HPC programs, and enterprise AI factories, the EX4000 is notable for delivering exaflop-class FP8 training density in a footprint that scales from a single cabinet to multi-thousand-node installations. Its tight integration of NVIDIA's GB200 NVL72 reference topology with HPE's decades of Cray supercomputing chassis engineering produces a system that achieves both memory bandwidth and interconnect bandwidth at a scale previously confined to purpose-built supercomputer deployments, making frontier-class AI model training accessible within a commercially orderable platform.

Ideal for

• Large language model (LLM) pre-training at trillion-parameter scale using FP8 and BF16 mixed-precision Tensor Core operations across the full NVLink-connected GPU pool
• Computational fluid dynamics (CFD) and finite element analysis (FEA) simulations for aerospace, automotive, and energy sector engineering requiring extreme parallel floating-point throughput
• Climate and weather modeling at global resolution, leveraging GPU-accelerated numerical solvers and the high-bandwidth Slingshot fabric for distributed memory communication
• Genomics and drug discovery pipelines running GPU-accelerated molecular dynamics (GROMACS, NAMD, AlphaFold) at population-scale dataset volumes
• Generative AI inference serving for large multimodal models requiring NVLink-unified GPU memory pools to hold multi-hundred-billion parameter models in active memory
• Government and defense high-performance computing workloads including signals intelligence processing, Monte Carlo simulations, and classified AI model development on air-gapped supercomputer fabrics

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.