NVIDIA GH200 480GB NVL32 Grace Hopper Superchip for HPE Cray XD670

The NVIDIA GH200 480GB NVL32 Grace Hopper Superchip for HPE Cray XD670 (P57602-B21) is a purpose-built accelerated computing module that unifies an NVIDIA Grace CPU with an NVIDIA Hopper GPU on a single package, connected via NVLink-C2C with 900 GB/s of chip-to-chip bandwidth. The GH200 Superchip delivers 96GB of HBM3 GPU memory paired with 480GB of LPDDR5X CPU memory in the NVL32 configuration, enabling exceptionally large model footprints to reside entirely in high-bandwidth memory across a 32-node NVLink domain without reliance on slower PCIe interconnects.

Designed natively for the HPE Cray XD670 platform, this module integrates seamlessly into HPE's liquid-cooled, high-density accelerated compute architecture. The Hopper GPU die incorporates fourth-generation Tensor Cores with FP8 precision support, a Transformer Engine for dynamic mixed-precision inference and training, and Confidential Computing capabilities via hardware-level TEE isolation. The NVL32 multi-node NVLink configuration allows up to 32 GH200 Superchips to share a unified 57.6 TB/s aggregate NVLink bandwidth fabric, enabling tight coupling of GPU and CPU memory across nodes at scales not achievable with conventional interconnects.

This accelerator is engineered for the most demanding generative AI, large language model training and inference, scientific simulation, and high-performance computing workloads in enterprise data centers. Its combination of massive coherent memory capacity, extreme interconnect bandwidth, and Hopper-generation compute throughput positions the GH200 NVL32 as a foundational building block for next-generation AI infrastructure deployed across hyperscale, national laboratory, and enterprise research environments.

Ideal for

• Large language model (LLM) pre-training and fine-tuning requiring hundreds of gigabytes of unified high-bandwidth memory across tightly coupled GPU nodes
• Generative AI inference serving for frontier models where low-latency, high-throughput token generation demands maximum memory bandwidth and capacity
• High-performance scientific computing and simulation workloads such as molecular dynamics, climate modeling, and computational fluid dynamics that benefit from coherent CPU-GPU memory access
• Retrieval-augmented generation (RAG) pipelines and vector database workloads that require fast, large-scale in-memory data access across accelerated compute nodes
• Multi-node deep learning model parallelism using NVLink-connected GPU clusters to eliminate PCIe bandwidth bottlenecks in distributed training jobs
• Confidential AI computing scenarios requiring hardware-enforced memory encryption and trusted execution environments for sensitive enterprise data processing

Technical specifications

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