HPE Cray XD670
$150,000.00
HPE Cray XD670 — The HPE Cray XD670 is a 5U, single-node GPU-accelerated server engineered for large AI training and HPC workloads. HPE’s current QuickSpecs describe a dual-socket CPU node with eight NVIDIA H100 or H200 Tensor Core SXM5 GPUs, support for air cooling or direct liquid cooling, and up to 32 DDR5 DIMM slots. It should be purchased as a fully engineered rack-scale system, not as a generic eight-GPU server. This page is structured for buyers comparing HPE Cray XD670 specifications, price, compatibility and the exact configuration required before a B2B order.
Explore Related
Related products, categories and research.
HPE Cray XD670: premium buying summary
The HPE Cray XD670 is a 5U, single-node GPU-accelerated server engineered for large AI training and HPC workloads. HPE’s current QuickSpecs describe a dual-socket CPU node with eight NVIDIA H100 or H200 Tensor Core SXM5 GPUs, support for air cooling or direct liquid cooling, and up to 32 DDR5 DIMM slots. It should be purchased as a fully engineered rack-scale system, not as a generic eight-GPU server.
Key HPE Cray XD670 specification highlights
- 5U single-node server chassis.
- Eight NVIDIA H100 or H200 Tensor Core SXM5 GPUs; PCIe GPUs are not supported in the XD670 GPU baseboard architecture.
- H100 configurations use supported 4th Gen Intel Xeon Scalable processors; H200 configurations use supported 5th Gen Intel Xeon Scalable processors in the current QuickSpecs.
- Up to 32 DDR5 DIMM slots, with HPE listing up to 4 TB host memory in the current configuration table.
- Air-cooled and direct-liquid-cooled infrastructure options.
- High-speed networking/fabric and storage options are configuration-dependent.
HPE Cray XD670 product overview
The XD670 is purpose-built around tightly coupled SXM GPU compute. That is different from a conventional rack server where administrators choose a few PCIe accelerator cards. In the XD670, the GPU subsystem, CPU generation, host memory population, networking fabric, power and cooling have to be qualified together. HPE positions the system for large-model AI training, deep learning and advanced HPC simulation where eight-way GPU scale and fast GPU interconnect are central to the workload.
HPE Cray XD670 price, cost and quotation considerations
HPE Cray XD670 price comparisons must be configuration-specific. A system with eight H100 GPUs is not the same commercial product as an eight-H200 configuration, and direct liquid cooling can change rack, manifold and facility requirements. The quote should therefore list GPU generation/capacity, both CPUs, host-memory population, local storage, fabric adapters, cooling type, rack/PDU requirements, support level and deployment services.
Core features and capabilities
- Eight SXM5 GPU architecture for large-scale AI/HPC compute.
- Current H100 and H200 accelerator options.
- Dual Intel Xeon host architecture matched to the GPU generation in current HPE QuickSpecs.
- 32 DDR5 DIMM slots with high-memory configurations.
- Direct liquid cooling option for dense thermal loads.
- Support for high-speed fabrics/network adapters and integrated storage options.
- HPE service, remote-support and enterprise lifecycle tooling are available according to the purchased service agreement.
Detailed HPE Cray XD670 technical specifications
| Chassis | 5U, single node |
|---|---|
| GPU support | 8 × NVIDIA H100 80GB or H200 141GB Tensor Core SXM5 configurations in current HPE QuickSpecs |
| PCIe GPU support | Not supported for the primary GPU subsystem |
| CPU support | H100: supported 4th Gen Intel Xeon Scalable; H200: supported 5th Gen Intel Xeon Scalable |
| CPU count | 2 processors |
| Memory slots | 32 DDR5 RDIMM slots; 8 channels per processor, 2 DIMMs per channel |
| Maximum host memory | Up to 4 TB in current HPE QuickSpecs |
| Memory speed | Current configurations include DDR5-4800 and DDR5-5600 options; population rules affect achievable speed |
| Dimensions | 22.23 cm H × 44.8 cm W × 94.5 cm D |
| Maximum listed weight | 240 lb / 108.9 kg |
| Cooling | Air-cooled or direct-liquid-cooled options depending configuration |
| Networking | Multiple high-speed Ethernet/fabric options; exact adapter count and fabric depend on CTO configuration |
| Management/support | HPE service and remote-support options according to warranty or contractual support agreement |
Performance and use-case information
The main performance choice is H100 versus H200. H200 provides substantially more GPU memory per accelerator and is paired with current 5th Gen Xeon configurations in HPE’s QuickSpecs, while H100 configurations use supported 4th Gen Xeon processors. The benefit of either eight-GPU system depends on software parallelism, interconnect, data pipeline, storage and fabric. A buyer should avoid quoting expected model-training speed from GPU count alone.
Compatibility, software, connectivity and site requirements
XD670 compatibility starts at the rack and facility level. Confirm rack type, cooling architecture, CDU/manifold requirements for DLC, power distribution, network fabric, operating system, GPU software stack, storage path and service access. HPE notes specific rack hardware for certain direct-liquid-cooled configurations. The system is not a drop-in replacement for a standard 1U/2U server and requires experienced data-center deployment planning.
Applications
- Large-language-model training
- Deep-learning model training and fine-tuning
- HPC simulation
- Scientific computing
- Cryo-EM reconstruction and other GPU-accelerated research
- Large-scale accelerated analytics
- Enterprise AI infrastructure
Comparison with closely related models
Within the HPE Cray XD670 family, H100 versus H200 is the most useful buyer comparison because both use the same 5U eight-GPU concept but target different memory and CPU generations. Compared with standard PCIe GPU servers, XD670 is more specialized around eight SXM GPUs and dense interconnect. Buyers that do not need eight-way SXM scale should compare a conventional PowerEdge or ProLiant platform rather than assuming XD670 is automatically the right AI server.
What is included?
There is no single universal XD670 package. The quotation must identify the CTO base server, exact eight-GPU module, two processors, memory kits, storage, networking/fabric adapters, cooling configuration, rack/PDU or DLC components, cables and HPE service. Installation and integration services should be stated separately.
Important limitations and buyer considerations
- Requires expert rack, power, network and cooling planning.
- PCIe GPUs are not supported as substitutes for the eight-GPU SXM subsystem.
- H100 and H200 configurations have different current CPU-generation requirements.
- The current HPE QuickSpecs show a minimum-weight line with inconsistent lb/kg units; use the exact configured system BOM and shipping documentation for freight planning rather than relying on that minimum line.
- GPU performance is workload/software dependent.
- Warranty level and remote support depend on the specific HPE service agreement.
Who should buy this exact model?
The HPE Cray XD670 is appropriate for organizations that genuinely need eight tightly coupled H100/H200-class GPUs for model training or HPC. It is excessive for ordinary virtualization, web hosting or small inference workloads, where a standard rack server can be easier and cheaper to deploy.
Authoritative technical and manufacturer resources
For this server, validate the current HPE CTO bill of materials, rack/cooling architecture and support entitlement before placing an order.
HPE Cray XD670 technical quick answers
What GPUs does the HPE Cray XD670 support?
HPE’s current QuickSpecs list eight NVIDIA H100 or H200 Tensor Core SXM5 GPUs.
How large is the the XD670?
It is a 5U chassis about 22.23 cm high, 44.8 cm wide and 94.5 cm deep.
How much host memory can the the XD670 support?
The current HPE QuickSpecs list up to 4 TB of DDR5 host memory.
Does the XD670 support liquid cooling?
Yes. HPE lists direct-liquid-cooling configurations in addition to air-cooled options.
Detailed FAQs
Can I install PCIe GPUs in place of the H100/H200 SXM GPUs?
HPE states PCIe GPUs are not supported for the XD670 primary GPU configuration.
Which CPUs are used with H200?
Current HPE QuickSpecs list support for 5th Gen Intel Xeon Scalable processors with H200 configurations.
What should be confirmed before ordering?
GPU generation, CPU SKUs, memory, storage, fabric, cooling, rack/PDU, software, support, installation and exact system weight.
Is the XD670 suitable for a normal server room?
Only if the facility can meet its rack, power, thermal and service requirements; DLC configurations can require additional infrastructure.
How should warranty be verified?
HPE directs buyers to its warranty-check tool and service agreement; the quote should state the exact support level.
Relevant internal links
Request a B2B quotation
For an XD670 quotation, specify H100 or H200, host-memory target, fabric/network requirements, cooling method, rack destination and HPE service level so the CTO configuration can be priced correctly.
