HPE ProLiant DL585a Gen13

The HPE ProLiant DL585a Gen13 is a high-density, 10U dual-socket enterprise server engineered specifically for sustained AI workloads, retrieval-augmented generation (RAG), and agentic workflows. Built to house two 6th Gen AMD EPYC 9006 "Venice" processors alongside up to eight double-wide PCIe accelerator cards, this system bridges the gap between traditional enterprise virtualization and dedicated scale-up AI supercomputers. It brings data-center compute directly into standard 19-inch rack infrastructure without requiring proprietary OCP rack modifications or specialized liquid-to-air facility retrofits.

Where modern multi-node clusters often force teams into rigid, proprietary interconnect topologies, the DL585a Gen13 prioritizes modular PCIe Gen6 I/O and balanced CPU-GPU resource allocation. Enterprise deployments frequently falter when high-throughput inference engines outrun their host CPUs during pre-processing, context retrieval, and vector database queries. By pairing up to 512 physical CPU cores with massive system memory bandwidth and eight discrete accelerator slots, this chassis targets teams looking for the best hardware for running AI models locally within enterprise-controlled boundaries.

For infrastructure architects sourcing Unknown hardware for AI development across multi-vendor fleets, the platform provides a flexible hardware footprint. Rather than locking administrators into a single accelerator ecosystem, the DL585a Gen13 accepts double-wide PCIe accelerators from both Nvidia and AMD. This makes the HPE ProLiant DL585a Gen13 for AI a versatile host platform for enterprise clusters dedicated to high-volume token generation, parameter-efficient fine-tuning (PEFT), and multi-agent coordination.

The ProLiant DL585a Gen13 is built from the ground up to eliminate host-to-accelerator I/O bottlenecks. Deploying modern local models requires moving massive KV caches and model weights across PCIe lanes with minimal latency.

At the core of the system are two 6th Gen AMD EPYC 9006 SP7 processors delivering up to 256 physical cores each (512 cores and 1024 threads total). This host capacity is critical for running dense vector search indices, pre-tokenization pipelines, and agent runtime orchestration on the CPU side while GPUs remain focused on tensor operations.

• GPU Capacity: Up to 8 double-wide PCIe accelerator cards (supporting Nvidia or AMD hardware).
• PCIe Connectivity: 16-channel PCIe Gen6 architecture, offering double the per-lane bandwidth of PCIe Gen5. This guarantees that host-to-device offloading, context window loading, and inter-GPU data exchanges operate without interface saturation.
• Physical Expansion: Up to six front-accessible and two rear-accessible PCIe expansion slots allow flexible cabling and card arrangements.

Memory throughput is a primary limiter during autoregressive token generation. The DL585a Gen13 features 32 DIMM slots operating across 16 channels, supporting:

• Standard DDR5 memory at speeds up to 8000 MT/s.
• Multiplexer Combined R-DIMM (MRDIMM) modules reaching up to 12800 MT/s.

Using MRDIMM configurations provides system memory bandwidth that rivals low-tier accelerators. This capability allows high-speed CPU offloading for large parameter weights that exceed discrete accelerator VRAM.

Inter-node scaling and remote direct memory access (RDMA) over converged Ethernet (RoCE) are integrated directly into the chassis:

• Fabric: Integrated 8x QSFP fabric module with an Nvidia switchboard.
• Offload and SmartNICs: Native hardware support for Nvidia BlueField-3 and BlueField-4 Data Processing Units (DPUs), as well as ConnectX-8 and CX9 SuperNIC adapters for low-latency multi-node scaling.

Delivering power to eight high-TDP accelerators demands extensive electrical headroom:

• Dual 3.2kW power supplies dedicated solely to the system board and CPUs.
• Modular expansion bays holding up to 12 additional 3.2kW power supplies.
• Total power availability scales well beyond 40kW if fully populated, accommodating high-transient power draws from next-generation enterprise GPUs.

System administration is handled via HPE iLO 8, which includes automated self-encrypting drive (auto-SED) management and multi-party authorization to secure model weights and proprietary data stored at rest.

When outfitting the system with eight modern GPUs, the HPE ProLiant DL585a Gen13 VRAM for large language models depends on the installed cards. Populating the chassis with eight 80GB or 96GB PCIe accelerators yields 640GB to 768GB of unified VRAM. Populating it with enterprise-tier accelerators sporting larger memory footprints can push total GPU memory past 1TB.

The chassis easily acts as hardware for running 70B parameter models at massive concurrency. Models such as Llama 3.1 70B and Qwen 2.5 72B require roughly 140GB of VRAM in unquantized 16-bit precision:

• Distribution: A single 70B model fits comfortably across two 80GB cards, allowing the server to host up to four independent 70B inference instances simultaneously.
• Quantization Sweet Spot: Running FP8 or AWQ 4-bit quantizations allows a 70B model to fit onto a single accelerator. This yields massive parallel capacity, serving dozens of simultaneous agentic threads.
• Throughput: For an HPE ProLiant DL585a Gen13 local LLM stack serving Llama 3.1 70B across tensor-parallel pipelines, expect aggregate throughput exceeding 120 to 180 tokens per second per GPU pair under high batch sizes.

Hosting giant open-weight models like Llama 3.1 405B requires distributed tensor and pipeline parallelism:

• 16-bit (BF16): Requires over 810GB of VRAM. A full chassis of eight 96GB or 144GB PCIe GPUs can host this model locally without host RAM offloading.
• 8-bit (FP8) Quantization: Consumes approximately 420GB to 450GB of VRAM. An eight-GPU configuration running an FP8 build fits the weights with ample leftover memory allocated to 128k context KV caches.
• Token Generation: While cross-card PCIe Gen6 latency is slightly higher than dedicated SXM baseboards, the DL585a Gen13 tokens per second remain well within production serving thresholds (15 to 30 tokens/sec across an eight-way tensor-parallel pipeline for interactive queries).

Models such as Mixtral 8x22B and DeepSeek-R1 (671B MoE) benefit significantly from the system's vast memory capacity:

• DeepSeek-R1 running in 4-bit or dynamic quantized variants fits entirely across eight high-capacity PCIe cards.
• The 512 AMD EPYC Venice CPU cores efficiently manage token routing and expert activation logic when orchestrating dense multi-GPU MoE inference.

Processing vision-language models like Qwen 2.5-VL or processing 128k-to-1M token contexts requires vast memory allocations for KV cache retention. The high system memory bandwidth supported by MRDIMM (12800 MT/s) makes CPU-side KV cache swapping viable when requests exceed physical GPU VRAM.

The HPE ProLiant DL585a Gen13 is strictly an enterprise-grade and research-tier system. It is engineered for:

• Enterprise Agentic Workflows: Ideal as the best hardware for local AI agents 2026 demands, where systems must run local vector search, evaluate tool calls, parse multi-step reasoning, and generate model responses simultaneously without leaving an on-prem data center.
• Production Inference Clusters: IT departments centralizing inference endpoints for thousands of internal users querying proprietary code bases, documentation, and sensitive financial records.
• RAG and Parameter-Efficient Fine-Tuning: Perfect for teams running continuous LoRA and QLoRA fine-tuning jobs alongside production vector databases, avoiding the external API costs and data security concerns of cloud providers.
• High-Density Virtualization: Environments running mixed compute workloads where administrators slice GPU resources using vGPU profiles to serve both traditional enterprise analytics and dedicated local LLM development sandboxes.

This machine is not intended for single-seat developers or hobbyist desk setups. The 10U physical form factor, massive acoustic footprint, and multi-kilowatt power supply demands require standard data-center rack installations, high-voltage PDUs, and dedicated cold-aisle cooling.

When evaluating the HPE ProLiant DL585a Gen13 AI inference performance, architects typically compare it against dedicated SXM-based servers and liquid-cooled OCP hardware.

The Dell PowerEdge XE9680 is a popular 8-GPU server based on dedicated HGX/SXM modules or OAM form factors.

• Interconnect: The XE9680 uses NVLink or Infinity Fabric backplanes offering up to 900 GB/s to 1.8 TB/s of GPU-to-GPU bandwidth, which gives it an advantage in pre-training large models from scratch.
• Flexibility: The DL585a Gen13 counters with standard double-wide PCIe Gen6 slots. This allows teams to hot-swap cards, mix hardware revisions, or transition between Nvidia and AMD accelerators without replacing the baseboard.
• CPU Density: The DL585a Gen13 provides up to 512 AMD Venice cores, doubling the CPU core ceiling of many competing host systems to better balance data prep and agent execution.

Within HPE's own Gen13 lineup, the XD245 targets ultra-dense hyperscale environments.

• Form Factor: The XD245 is an Open Rack (21-inch) liquid-cooled design built for massive scale-out facilities.
• Deployment Ease: The DL585a Gen13 uses a standard 19-inch 10U rack footprint and air cooling, making it far simpler to deploy inside existing enterprise data centers that lack facility-level liquid loops.

For enterprise infrastructure teams prioritizing open accelerator selection, massive CPU compute, and standard rack deployment over closed, proprietary baseboards, the DL585a Gen13 stands as a top-tier platform for production AI workloads.