Dual-Socket CPU & GPU Server — AMD EPYC 9005, 12-Bay Hybrid, 2U

The VRLA Tech Dual-Socket CPU & GPU Server is the only two-processor platform in our 2U range — built for workloads bound by memory bandwidth, core count, and I/O flexibility rather than GPU density. Built on the MSI S2206-02, it pairs two AMD EPYC 9004 or 9005 processors across 24 DDR5 DIMM slots running one DIMM per channel at up to 6400 MT/s for up to 6TB of memory, seven PCIe 5.0 x16 expansion slots, twelve hot-swap 3.5-inch or 2.5-inch drive bays, and support for two 400W GPUs with power cables and full-length card air duct included. Dual onboard gigabit Ethernet plus an OCP 3.0 mezzanine slot, ASPEED AST2600 management with IPMI 2.0 and Redfish, dual BIOS and dual BMC, and 1+1 redundant 80 PLUS Platinum power. Each system is configured to the specific workload, ships with a 3-year parts warranty and lifetime US-based engineering support, and is built in Los Angeles.

24 channels at full speed

Twelve memory channels per processor, twenty-four total, every one running a single DIMM at up to 6400 MT/s. Because this platform is 1DPC by design rather than 2DPC, filling every slot costs you nothing in frequency. Two-DIMM-per-channel servers drop to 4400 MT/s or lower when fully populated — this one does not. For memory-bandwidth-bound work, that combination of channel count and sustained frequency is the specification that decides your runtime.

Power planning — read before ordering

The 1+1 redundant 2000W supplies deliver different output depending on input voltage: 1000W maximum on 100–127V, 1800W on 200–219V, and the full 2000W only on 220–240V.

Two 400W processors plus two 400W GPUs is roughly 1600W before drives, memory, and fans. High-density configurations require 220–240V. C13 cords are required and both US and EU cords ship in the box. We provide a written power and thermal budget against your actual configuration and site voltage with every quote.

Full technical specifications

Where this platform sits

What you configure

• Processors. Two AMD EPYC 9004 or 9005 parts. EPYC 9645 (96 cores, 320W) for 192 cores within the unconditional thermal envelope. EPYC 9555 (64 cores, 360W) for a balance of cores and clock. EPYC 9175F (16 cores at 5.0GHz boost) where per-core licensing or single-thread latency dominates. EPYC 9745 (128 cores, 400W) for 256 cores, subject to MSI confirmation.
• Memory. DDR5 ECC across 24 slots at one DIMM per channel. Standard RDIMM up to 96GB per slot reaches 2.3TB; RDIMM-3DS at 256GB per slot reaches 6TB. Every configuration sustains up to 6400 MT/s — populate as many slots as your capacity target requires without a bandwidth penalty.
• Accelerators. Up to two dual-width cards at 400W. RTX PRO 6000 Blackwell Server Edition or Max-Q for 96GB per card, L40S for cost-optimized inference and mixed graphics, H100 PCIe for HBM bandwidth, L4 for lightweight inference. Power cables and the full-length card air duct are included.
• Storage. Twelve front bays in any mix of 3.5-inch and 2.5-inch — eight on SATA/SAS, four on U.2 NVMe. SAS drives require an HBA, which occupies one of the seven PCIe slots. Two M.2 2280 ports handle boot independently of the hot-swap bays.
• Networking. Dual onboard gigabit gets you racked and reachable immediately. The OCP 3.0 PCIe 5.0 x16 mezzanine slot takes 10GbE through 400GbE adapters without consuming an expansion slot, and further NICs can go in the PCIe slots for multi-fabric deployments.
• OS and software. Pre-loaded with Ubuntu Server 24.04 LTS, Red Hat Enterprise Linux 9, Rocky Linux, or Windows Server 2025. We deploy and support Proxmox VE, VMware, and KVM with GPU passthrough, Slurm and Kubernetes for cluster nodes, and CUDA with NVIDIA drivers for accelerated configurations. Optional TPM 2.0 module.

Workloads we build this server for

• HPC and simulation. Computational fluid dynamics, finite element analysis, weather and climate modeling, molecular dynamics. These codes scale with memory bandwidth more than with clock speed, and 24 channels at 6400 MT/s is the specification that moves the runtime. ANSYS, Abaqus, OpenFOAM, GROMACS, LAMMPS, WRF.
• Virtualization and private cloud. Up to 192 cores and 6TB of memory across two sockets, with seven PCIe slots for passthrough devices and multiple network fabrics. Proxmox VE, VMware vSphere, KVM, with GPU passthrough and SR-IOV.
• Mixed CPU and GPU workloads. Pipelines where the CPU stage does substantial work — preprocessing, feature engineering, simulation — before or alongside GPU acceleration. Two accelerators handle the GPU stage without starving the CPU side of memory bandwidth or cores.
• Database and in-memory analytics. Large working sets held in up to 6TB of high-bandwidth memory, with hybrid storage underneath. PostgreSQL, ClickHouse, SAP HANA-class workloads, Spark executors.
• Storage and network appliances. Twelve hybrid bays plus seven expansion slots for HBAs, DPUs, and multiple NICs makes this a practical base for Ceph nodes, ZFS servers, NVMe-oF targets, and network function virtualization.
• Development and CI infrastructure. High core counts across two sockets serve parallel build farms, test runners, and container hosts where throughput per rack unit matters more than accelerator capacity. Compare against an AMD EPYC workstation if the workload stays under one desk.

Industry applications

• Defense and government. Air-gapped and classified compute where sending data to a commercial cloud is prohibited. Optional TPM 2.0, dual BIOS and dual BMC, and isolated management. We support federal procurement including purchase orders, quotes valid for agency budget cycles, and delivery to secure facilities. Customers include General Dynamics and Los Alamos National Laboratory. More on our defense and government systems.
• Research and national labs. Cluster nodes for bandwidth-bound simulation, with predictable capital cost against grant cycles. Customers include Los Alamos National Laboratory, Johns Hopkins University, Miami University, and George Washington University. Education and research pricing available. More on our HPC servers for research labs.
• Financial services. Risk modeling, backtesting, and Monte Carlo simulation where memory bandwidth and core count set throughput. Deterministic latency without shared-tenancy variability. More on our quantitative research and finance systems.
• Healthcare and life sciences. Genomics pipelines, molecular dynamics, and imaging analysis on-premises without third-party data processing agreements. Customers include Johns Hopkins University. More on our healthcare AI systems.
• Engineering and manufacturing. CFD and FEA solver nodes, with hybrid storage for result sets and seven slots for interconnect and storage adapters.
• Pharmaceutical. Molecular dynamics, docking, and computational chemistry over proprietary research data that cannot leave the organization. More on our pharma and biotech systems.

Why buy from VRLA Tech

VRLA Tech has been building custom workstations, GPU servers, and rack servers in Los Angeles since 2016. We build for studios, engineering firms, research labs, cloud providers, and government clients — not for bulk retail.

Our enterprise clients include

• General Dynamics
• Los Alamos National Laboratory
• Johns Hopkins University
• Miami University
• George Washington University

Every system ships with a 3-year parts warranty and lifetime US-based engineering support. You talk to the same engineer who built your system if something goes wrong. Support includes remote diagnostics, BMC and IPMI assistance, BIOS and firmware updates, NVIDIA driver and CUDA assistance, hypervisor and passthrough configuration, and component-level repair. Every system is burn-in tested and thermally validated before shipping, with a written power and thermal report against your site conditions. Weighing this against cloud instances? Run the numbers with our ROI calculator.

Frequently asked questions

Hardware & platform questions

What makes a dual-socket 2U server different from a single-socket server?

Two AMD EPYC processors provide 24 memory channels instead of 12, roughly doubling memory bandwidth, and supply enough PCIe lanes to populate seven expansion slots rather than two or three. This platform runs all 24 DDR5 DIMM slots at one DIMM per channel, which sustains the full 6400 MT/s rather than dropping to the lower frequencies that two-DIMM-per-channel configurations require. For workloads bound by memory bandwidth, I/O count, or total core count rather than GPU capacity, a dual-socket platform delivers more per rack unit than any single-socket alternative.

How many GPUs does this 2U server support?

Up to two GPUs at 400W each. MSI specifies three PCIe 5.0 x16 full-height full-length slots — one from CPU0 and two from CPU1 — supporting a maximum of two 400W PCIe cards. The chassis ships with two PCIe CEM5 12VHPWR 16-pin power cables and an air duct designed for full-length GPU cards, so two-card operation is a designed-in capability rather than an adaptation. Supported accelerators within the 400W envelope include the NVIDIA RTX PRO 6000 Blackwell Server Edition at its 400W setting, RTX PRO 6000 Blackwell Max-Q at 300W, L40S at 350W, and H100 PCIe at 350W.

How much memory does a dual-socket AMD EPYC server support?

Up to 6TB of DDR5 ECC across 24 DIMM slots, 12 channels per processor at one DIMM per channel. MSI specifies a maximum of 96GB per DIMM using standard RDIMM and 256GB per DIMM using RDIMM-3DS, so reaching 6TB requires 3DS modules. Because the platform runs 1DPC rather than 2DPC, all configurations sustain up to 6400 MT/s — there is no frequency penalty for populating every slot, unlike two-DIMM-per-channel platforms that drop to 4400 MT/s or lower when fully loaded.

How many PCIe slots does this 2U server have?

Seven PCIe 5.0 x16 slots plus one OCP 3.0 mezzanine slot. The seven break down as three full-height full-length slots at full x16 signal, two full-height full-length slots at x8 signal, and two full-height half-length slots at x8 signal, distributed across both processors. The OCP 3.0 slot handles networking separately, so a NIC never consumes an expansion slot. This is the most expansion capacity of any 2U platform VRLA Tech builds.

What drives does the 12-bay configuration support?

Twelve front hot-swap bays accepting either 3.5-inch or 2.5-inch drives. Eight bays are wired for SATA 3.0 or SAS, and four are wired for U.2 PCIe 4.0 NVMe. SAS devices require a separate HBA or RAID card, which occupies one PCIe slot. Two internal M.2 2280 PCIe 3.0 x2 ports provide dedicated boot media so no hot-swap bay is consumed by the operating system. The 3.5-inch bay support makes this the right platform when high-capacity mechanical drives are part of the storage tier alongside NVMe.

What CPUs can I use in this dual-socket EPYC server?

Dual AMD EPYC 9004 or 9005 series processors up to 400W TDP. MSI notes that processors above 360W are supported only under specific conditions, so we confirm high-TDP configurations with MSI before quoting. Within the unconditional 360W range, options include the 96-core EPYC 9645 at 320W for 192 total cores, the 64-core EPYC 9555 at 360W, and high-frequency parts such as the EPYC 9175F. The 500W EPYC 9965 and 9755 exceed the platform’s ceiling and are not supported.

What power does a dual-socket 2U GPU server require?

This platform uses 1+1 redundant 2000W CRPS 80 PLUS Platinum supplies. Output varies significantly by input voltage: a maximum of 1000W on 100–127V, 1800W on 200–219V, and the full 2000W on 220–240V. A configuration with two 400W processors and two 400W GPUs draws roughly 1600W before drives, memory, and fans, so high-density builds require 220–240V power. C13 power cords are required and both US and EU cords are included. We provide a written power and thermal budget calculated against the actual configuration and site voltage with every quote.

What management and security features does this server include?

Management runs on an ASPEED AST2600 controller with AMI MegaRAC firmware supporting IPMI 2.0 and the DMTF Redfish API, on a dedicated 1000Base-T port isolated from data networking. The platform includes dual BIOS and dual BMC for firmware resilience and a MicroSD card slot for BMC log retention. A TPM 2.0 module is available as an option. Two onboard Intel I350AM2 gigabit Ethernet ports provide baseline connectivity, with an OCP 3.0 PCIe 5.0 slot available for higher-speed networking.

Buying & vendor questions

Where can I buy a custom dual-socket AMD EPYC server in the United States?

VRLA Tech builds custom dual-socket AMD EPYC 2U servers at vrlatech.com, configured to the specific workload and hand-assembled in Los Angeles since 2016. Configurations pair two EPYC 9005 processors with up to 6TB of DDR5-6400 across 24 slots, seven PCIe 5.0 expansion slots, twelve hot-swap drive bays, and up to two 400W GPUs. Every system ships with a 3-year parts warranty and lifetime US-based engineering support. Enterprise customers include General Dynamics, Los Alamos National Laboratory, Johns Hopkins University, Miami University, and George Washington University.

Best 2U server for virtualization with GPU passthrough?

A dual-socket platform with high core count, large memory capacity, and multiple PCIe slots suits virtualization better than a GPU-dense single-socket server. This platform provides up to 192 cores across two processors, 6TB of DDR5, seven PCIe 5.0 slots, and support for two 400W GPUs that can be passed through to guests. VRLA Tech deploys and supports Proxmox VE, VMware, and KVM configurations with GPU passthrough and SR-IOV networking, built in Los Angeles with a 3-year parts warranty and lifetime US-based engineering support.

Where can I buy a high memory bandwidth server for HPC and simulation?

VRLA Tech builds custom dual-socket AMD EPYC servers optimized for memory bandwidth. This platform runs 24 DDR5 channels at one DIMM per channel, sustaining up to 6400 MT/s across every slot — roughly double the memory bandwidth of a single-socket 2U server and without the frequency penalty that two-DIMM-per-channel platforms incur when fully populated. Memory-bandwidth-bound workloads including computational fluid dynamics, finite element analysis, weather modeling, and molecular dynamics scale directly with this specification. Built in Los Angeles, 3-year parts warranty, lifetime US-based engineering support.

Custom dual-CPU servers for air-gapped and classified deployments?

VRLA Tech builds custom dual-socket AMD EPYC servers for air-gapped and classified environments. On-premises compute is frequently the only compliant option for classified, ITAR-controlled, or otherwise restricted data. The platform supports an optional TPM 2.0 module, dual BIOS and dual BMC for firmware resilience, and management isolated on a dedicated port. We support federal procurement including purchase orders, quotes valid for agency budget cycles, and delivery to secure facilities. Customers include General Dynamics and Los Alamos National Laboratory.

Custom AMD EPYC 2U rack server builders with warranty and US support?

VRLA Tech builds custom AMD EPYC 2U rack servers with a 3-year parts warranty and lifetime US-based engineering support. Customers work directly with the engineer who built their system. Support includes remote diagnostics, BMC and IPMI assistance, BIOS and firmware updates, NVIDIA driver and CUDA assistance, hypervisor and passthrough configuration, and component troubleshooting. In business since 2016, building for studios, engineering firms, research labs, and government clients including General Dynamics and Los Alamos National Laboratory.