Description
Deliver leadership-class AI training and deep learning performance with the HPE Cray XD670, a dense 5U GPU server built on the NVIDIA HGX H100 platform. Equipped with 8 NVIDIA H100 SXM5 GPUs — each with 80GB of HBM3 memory — dual Intel Xeon Platinum 8462Y+ processors (32-core, 2.80 GHz, 300W TDP), and 2TB of DDR5 ECC memory, the XD670 is purpose-built for large-scale model training, LLM fine-tuning, generative AI, and advanced HPC simulation. With 640GB of total GPU memory and over 24 TB/s of aggregate HBM3 bandwidth across all eight accelerators, this system handles the most parameter-intensive workloads without compromise.
Each of the eight H100 SXM5 GPUs communicates at 900 GB/s bidirectional bandwidth via 4th-generation NVLink and NVSwitch, delivering 3.6 TB/s of total GPU-to-GPU interconnect throughput within a single node. For multi-node scaling, the XD670 is configured with 8 NVIDIA ConnectX-7 single-port OSFP adapters, each providing 400 Gb/s InfiniBand NDR or 400GbE connectivity — a total of 3.2 Tb/s of fabric bandwidth per server. Two NVIDIA BlueField-2 DPUs provide dedicated 2×25 Gb/s Ethernet ports for storage networking, infrastructure offload, and zero-trust security acceleration. An additional pair of 10GbE Intel X710-AT2 management ports round out the networking configuration.
This configuration ships without internal storage, designed for deployment with external high-performance storage systems such as Pure Storage FlashBlade or FlashArray for maximum flexibility in tiered and shared-storage architectures. Power is supplied by 6x 3000W Delta 80 PLUS Titanium power supplies in a redundant configuration, and the system is air-cooled for broad datacenter compatibility. Backed by HPE’s 3-year limited warranty with return-to-depot service, the XD670 is ready for immediate deployment in production AI factories, neocloud environments, and enterprise HPC clusters.
