Skip to content
Uniqcli

HPE ProLiant Compute XD685 vs HPE Cray XD670: Choosing the Right 8-GPU AI Server

The HPE ProLiant Compute XD685 and HPE Cray XD670 are both 8-way GPU servers within HPE's accelerated compute lineup, yet they come from two different design philosophies. The XD685 is an enterprise AI node built around 5th Gen AMD EPYC processors and AMD Instinct MI325X/MI300X accelerators, while the Cray XD670 carries HPE's Cray HPC lineage with dual Intel Xeon CPUs and the NVIDIA HGX H100/H200 platform. This comparison breaks down GPU architecture, CPU choice, memory, cooling, high-speed fabric, and procurement so AI, research, and federal buyers can match the right GPU node to the workload.

The short answer

Choose the HPE ProLiant Compute XD685 when you want an AMD-centric enterprise AI node with eight Instinct MI325X or MI300X accelerators, huge HBM capacity, and tight AMD EPYC/ROCm pairing for large language model training and inference at scale. Choose the HPE Cray XD670 when you are standardizing on the NVIDIA HGX H100/H200 platform and CUDA ecosystem, want Cray HPC-grade fabric flexibility (Slingshot 11, InfiniBand NDR, Ethernet), and need a GPU node that slots cleanly into supercomputing and deep-learning clusters. Both are 8-GPU, dual-socket, liquid-cooling-capable nodes, so the decision usually comes down to AMD Instinct vs NVIDIA HGX, the surrounding software stack, and your fabric standard.

HPE ProLiant Compute XD685 vs HPE Cray XD670, head to head

HPE ProLiant Compute XD685
HPE Cray XD670
GPU architecture
8x AMD Instinct MI325X or MI300X (OAM) with large HBM capacity
8x NVIDIA HGX H100 or H200 SXM5 Tensor Core GPUs
CPU platform
2x 5th Gen AMD EPYC 9005, up to 96 cores each, up to 5.0 GHzadvantage
2x Intel Xeon (4th Gen for H100, 5th Gen for H200)
Memory capacity
Up to 6.1 TB DDR5 across 24 DIMM slots
DDR5-5600 across up to 32 DIMM slots
High-speed fabric
GPU-direct networking for scale-out AI clusters
8x PCIe Gen5 slots for Slingshot 11, InfiniBand NDR or Ethernetadvantage
Software ecosystem
AMD ROCm, optimized for open AI frameworks on Instinct
NVIDIA CUDA, cuDNN, NeMo and the broad HGX software stackadvantage
Design heritage
Enterprise AI node engineered for fast cluster deployment
Cray supercomputing lineage tuned for HPC and large-model AI
Cooling
5U direct liquid cooling or 6U air, 6x 3000W Titanium PSUs
5U chassis with direct liquid cooling or air-cooled options
Federal / TAA
TAA-compliant configurations available for federal buyers
TAA-compliant configurations available for federal buyers

Specifications side by side

HPE ProLiant Compute XD685
HPE Cray XD670
GPU count
8 accelerators
8 GPUs
GPU options
AMD Instinct MI325X or MI300X
NVIDIA HGX H100 or H200 SXM5
CPU
2x 5th Gen AMD EPYC 9005 series
2x Intel Xeon Scalable (4th Gen for H100, 5th Gen for H200)
Max CPU cores
Up to 96 cores per socket
Depends on selected Xeon SKU
Memory
Up to 6.1 TB DDR5, 24 DIMM slots
DDR5-5600, up to 32 DIMM slots
Form factor
5U direct liquid cooled / 6U air cooled
5U chassis
Cooling
Air or direct liquid cooling
Air or direct liquid cooling (CoolIT DLC)
Power
6x 3000W @54V Titanium Plus PSUs (N+1)
Redundant high-wattage PSUs for 8-GPU load
Fabric / expansion
GPU-direct networking for scale-out AI
8x PCIe Gen5 HHHL slots: Slingshot 11, InfiniBand NDR, Ethernet
Primary software stack
AMD ROCm
NVIDIA CUDA / HGX stack
Design lineage
HPE ProLiant Compute enterprise AI
HPE Cray supercomputing
Primary workloads
LLM training, NLP, multimodal AI clusters
Large-model AI training and deep learning / HPC

Where HPE ProLiant Compute XD685 wins

  • Eight AMD Instinct MI325X/MI300X accelerators with very large aggregate HBM for memory-bound LLM and inference workloads
  • 5th Gen AMD EPYC 9005 CPUs deliver up to 96 cores per socket and up to 5.0 GHz for strong host-side throughput
  • Up to 6.1 TB of DDR5 system memory supports large in-memory datasets and data preprocessing
  • Engineered as an enterprise AI node for fast deployment of secure, energy-efficient AI clusters
  • Air or direct liquid cooling with 54V Titanium Plus power for dense, efficient racks

Where HPE Cray XD670 wins

  • Eight NVIDIA HGX H100/H200 SXM5 GPUs tap the mature CUDA and NVIDIA AI software ecosystem
  • Cray supercomputing lineage tuned for large-model AI training and HPC at cluster scale
  • 8x PCIe Gen5 slots support Slingshot 11, InfiniBand NDR and Ethernet for fabric flexibility
  • NVLink/NVSwitch within the HGX baseboard enables high-bandwidth GPU-to-GPU communication
  • Direct liquid cooling option (including CoolIT) improves power efficiency for sustained training

Which one should you buy?

Standardizing a new cluster on the NVIDIA HGX and CUDA ecosystem

Pick HPE Cray XD670. With 8x NVIDIA HGX H100/H200 SXM5 GPUs and the full CUDA software stack, the XD670 drops into NVIDIA-centric AI and HPC environments and reuses existing CUDA tooling and skills.

Memory-bound LLM training and inference favoring large HBM per GPU

Pick HPE ProLiant Compute XD685. Eight AMD Instinct MI325X/MI300X accelerators offer very high HBM capacity per GPU, which helps fit larger models and batch sizes for training and inference on an AMD ROCm stack.

Supercomputing or research cluster needing Slingshot or InfiniBand NDR fabric

Pick HPE Cray XD670. Its 8x PCIe Gen5 slots support Slingshot 11, InfiniBand NDR and Ethernet, and the Cray lineage is built for high-speed fabric scale-out in HPC and large AI clusters.

AMD EPYC-standardized enterprise rolling out AI clusters quickly

Pick HPE ProLiant Compute XD685. Dual 5th Gen AMD EPYC 9005 CPUs and Instinct GPUs keep the platform on a single AMD architecture, simplifying procurement and tuning for shops already invested in EPYC.

Mixed AI and traditional HPC simulation workloads in one node

Pick HPE Cray XD670. The Cray-derived design and NVIDIA HGX platform are well suited to both deep-learning training and classic HPC simulation where CUDA-accelerated libraries dominate.

Frequently asked

What is the main difference between the HPE ProLiant Compute XD685 and the Cray XD670?

The XD685 is an enterprise AI node built around 5th Gen AMD EPYC CPUs and eight AMD Instinct MI325X/MI300X accelerators on the ROCm stack, while the Cray XD670 carries HPE's Cray HPC lineage with dual Intel Xeon CPUs and eight NVIDIA HGX H100/H200 SXM5 GPUs on the CUDA stack. The core decision is AMD Instinct versus NVIDIA HGX and the software ecosystem around each.

Which server is better for large language model training?

Both are strong 8-GPU LLM training nodes. The XD685 is attractive for memory-bound models thanks to the large HBM capacity of AMD Instinct MI325X/MI300X and up to 6.1 TB of DDR5. The Cray XD670 is the natural choice when your training pipeline is built on NVIDIA CUDA, NeMo, and the broader NVIDIA AI ecosystem, or when you need Slingshot/InfiniBand fabric at cluster scale.

What GPUs do the XD685 and XD670 support?

The HPE ProLiant Compute XD685 supports eight AMD Instinct MI325X or MI300X accelerators. The HPE Cray XD670 supports eight NVIDIA HGX H100 or H200 SXM5 Tensor Core GPUs. They are both 8-way GPU servers but on different accelerator architectures.

What CPUs do these servers use?

The XD685 uses two 5th Gen AMD EPYC 9005 series processors with up to 96 cores per socket and clocks up to 5.0 GHz. The Cray XD670 uses two Intel Xeon Scalable processors, typically 4th Gen Sapphire Rapids for H100 configurations and 5th Gen Emerald Rapids for H200 configurations.

Do both servers offer liquid cooling?

Yes. The XD685 ships as a 5U direct liquid cooled or 6U air cooled system, and the Cray XD670 offers air-cooled and direct liquid cooled options (including CoolIT DLC). Direct liquid cooling improves power efficiency and rack density for sustained 8-GPU AI training workloads on either platform.

Which server is better for HPC versus enterprise AI?

The Cray XD670 leans toward HPC and large-model AI thanks to its Cray supercomputing lineage and fabric flexibility (Slingshot 11, InfiniBand NDR, Ethernet). The XD685 is positioned as an enterprise AI node for quickly deploying secure, energy-efficient AI clusters on AMD Instinct and EPYC. Many environments run both depending on the workload and ecosystem.

How do the high-speed networking options compare?

The Cray XD670 provides eight PCIe Gen5 half-height, half-length slots that support Slingshot 11, InfiniBand NDR, and Ethernet for direct switchable connections between fabric, GPUs, NVMe, and CPUs. The XD685 is designed for GPU-direct networking in scale-out AI clusters; confirm the exact NIC and fabric options for your target topology when configuring.

Are the XD685 and Cray XD670 available for federal and TAA/GPC/SAP/FAR procurement?

Yes. As an authorized HPE reseller, we can source TAA-compliant HPE ProLiant Compute XD685 and HPE Cray XD670 configurations and quote them through federal and SLED vehicles such as GPC, SAP, and FAR, including HPE GreenLake consumption models where applicable. We can also help size GPU, CPU, memory, and fabric options to your AI or HPC cluster requirements.

Build your HPE bill of materials.

Send us the requirement, the project, or an existing quote to beat. We come back with a validated, TAA-compliant HPE configuration and a real price, often below list.

connect [at] getuniqcli.com · Chicago, IL