GPUmachines

RS700A-E13-RS12U Review: 1U Dual EPYC 9005 Server

A pair of EPYC 9005 processors, 24 memory channels and 12 front bays make the RS700A-E13-RS12U a dense CPU node. The useful question is whether its 1U thermal and I/O limits suit the intended service.

RS700A-E13-RS12U Review: 1U Dual EPYC 9005 Server

A 1U server with two 400 W-class processors leaves little room for lazy thermal design. The ASUS RS700A-E13-RS12U packs dual AMD EPYC 9005 sockets, 24 DIMM slots and twelve front drives into that space, so its appeal is density: a large amount of CPU, memory bandwidth and NVMe I/O per rack unit.

Density has a cost. High-power processors, fast network cards and a full NVMe front end create heat, fan noise and concentrated rack power. The server belongs in a data centre with controlled airflow. It is not an office storage box, nor is it a GPU server disguised by a few empty PCIe slots.

Executive Summary

The ASUS RS700A-E13-RS12U is a dual-socket 1U rack server for AMD EPYC 9005 processors. ASUS specifies 24 DDR5 DIMM slots, twelve 2.5-inch hot-swap bays, two PCIe 5.0 x16 FHHL slots, two OCP 3.0 x16 positions, two internal DIMM.2 storage modules and redundant 2000 W power supplies. Management is handled through ASUS's server-management stack and dedicated management interface.

The system suits CPU-dense virtualisation, analytics, software-defined storage, metadata services, databases and preprocessing nodes feeding larger GPU infrastructure. Twelve NVMe bays can provide a fast local or distributed-storage tier, while the two OCP positions give the node enough network choice to serve data rather than trap it inside the chassis.

Its strongest argument is rack efficiency. Buyers who need more drive capacity per node, several GPUs or low acoustic output should choose another form factor. A 2U or 4U storage server will be easier to cool and can expose more bays; a GPU server will have the power and slot geometry needed for accelerators.

Configure the ASUS RS700A-E13-RS12U through GPUMachines.

Key Specifications

| Area | Verified platform detail | | --- | --- | | Form factor | 1U rack server | | CPU platform | Dual AMD EPYC 9005 | | CPU sockets | 2 | | GPU support | Not designed as a GPU server | | Memory | 24 DDR5 RDIMM slots, 12 channels per CPU, up to 6000 MT/s at 1 DIMM per channel according to ASUS | | Storage | Twelve 2.5-inch hot-swap bays with NVMe and configuration-dependent SATA/SAS support, plus two DIMM.2 modules | | PCIe expansion | Two PCIe 5.0 x16 FHHL slots | | Networking | Two OCP 3.0 PCIe 5.0 x16 positions plus dedicated management LAN | | Power | 1+1 redundant 2000 W 80 PLUS Titanium or Platinum supplies, configuration-dependent | | Management | Dedicated management LAN and ASUS server-management controller | | Best-fit workloads | Virtualisation, databases, analytics, scale-out storage, metadata services and CPU-side AI data preparation |

SAS operation requires the appropriate HBA or RAID card, and drive-protocol support depends on the selected backplane configuration. GPUMachines should confirm the exact bay map before quoting drives or controllers.

Platform Highlights

  • Two EPYC 9005 processors in 1U maximise compute density. The platform can put a very high core count and 24 memory channels into one rack unit. That helps licensed capacity planning only when software and power budgets can use both sockets.
  • Twelve front bays give the node a practical data tier. A 12-NVMe arrangement favours throughput and low latency. Mixed NVMe, SATA or SAS configurations can suit boot, capacity and endurance requirements, but the backplane and controller must match.
  • Two OCP 3.0 positions are valuable in a storage role. One can serve the main data network while the other provides redundancy, a separate storage fabric or a second tenant path. OCP cards also avoid consuming the two conventional expansion slots.
  • Two PCIe slots leave room for storage control or specialist I/O. RAID, HBA, security and extra networking are possible without removing the OCP adapters. Slot width, airflow and cable routing still need checking.
  • Tool-free and modular service parts matter more in 1U. Dense fan bars, processors and cables leave little hand room. Serviceability reduces downtime, although component temperatures and firmware remain the larger operational concerns.

Our Technical View

RS700A-E13-RS12U is a compute server with a useful storage front end, not a storage array in its own right. Twelve NVMe devices can produce substantial aggregate bandwidth, but the system's value depends on software that distributes data, protects it and exposes it over a network. Local RAID can protect against a drive failure; it cannot protect against losing the node.

Dual-socket density can be attractive for virtualisation and CPU-heavy preprocessing. It also introduces NUMA behaviour. Each processor owns memory channels and PCIe paths, so software should keep threads, memory and devices on the same socket where possible. A database or storage daemon that constantly crosses the inter-socket link may fail to deliver the result implied by the component list.

In an AI environment, this server makes more sense beside GPU nodes than inside the GPU partition. It can run metadata, orchestration, data transforms, tokenisation, decompression or a software-defined storage service. Those jobs often need cores, memory bandwidth and network I/O but do not benefit from occupying an expensive accelerator slot.

The 1U form factor should be treated honestly. High-core CPUs and NVMe media produce heat in a shallow vertical space. Fans will run hard under load, and air-cooled racks need enough pressure and cold-aisle supply. A small deployment may save more money by using 2U nodes with lower fan power and easier service access.

Best-Fit Workloads

Software-defined storage node

Twelve flash devices, dual OCP networking and substantial CPU resources can support a Ceph, NVMe-oF or other distributed-storage role. The software's replication or erasure-coding policy determines usable capacity and failure behaviour. Three or more nodes are normally required before “scale-out” means anything operationally.

Metadata and control services

Parallel filesystems, schedulers, Kubernetes control planes and data catalogues need reliable CPU and low-latency storage, but not necessarily GPUs. The RS700A can host these services with appropriate clustering. Do not place every control function on one physical node simply because it has enough cores.

Virtualisation and private cloud

Dual EPYC sockets and 24 DIMMs provide a dense VM host. Local NVMe can hold active guests or cache a shared backend. NUMA-aware VM sizing matters; very large guests may span sockets and experience different memory latency.

Database and analytics

Relational databases, search indexes and analytical engines can use the memory channels and fast local media. Drive endurance, write amplification and recovery time are as important as headline read throughput.

CPU-side AI pipeline work

Dataset cleaning, tokenisation, media decode and feature extraction can run on CPU nodes before data reaches GPU servers. The value is measured by sustained output to the accelerator estate. A preprocessing server that reads quickly but cannot transmit data fast enough still starves the GPUs.

Who Should Consider It

Consider this model when rack units are scarce and the workload can genuinely use two EPYC processors, 24 memory channels and fast local drives. It suits experienced infrastructure teams building repeatable nodes for virtualisation, storage or data services.

The buyer should already know the intended drive protocol, network speed and data-protection method. “Twelve drives and lots of cores” is not a storage architecture.

Who Should Not Buy It

Teams needing bulk hard-drive capacity should choose a chassis with 3.5-inch bays and more physical slots. A 24-bay 2U server is better when flash capacity per node matters more than rack-unit density. Workloads needing GPU acceleration belong on a PCIe GPU or HGX platform.

Do not install this server in an office or acoustic rack without checking noise and cooling. A smaller single-socket system will be cheaper for services that cannot use the second CPU. Buyers planning only one node should also reconsider software-defined storage, because one server cannot provide node-level resilience.

Architecture Notes

The two processors form separate NUMA domains. Memory should be populated symmetrically, and storage or network interrupts should be handled by cores near the relevant PCIe device. Linux tools can report locality, but the application and scheduler must act on it.

ASUS provides 24 DIMM slots, which allows one DIMM per memory channel across both CPUs. That is a clean high-bandwidth population. Larger capacity may require different DIMM choices; confirm speed and rank restrictions for the selected EPYC CPUs.

The twelve front bays have multiple possible protocol maps. An all-NVMe layout removes the need for a SAS controller, while SATA or SAS requires the correct backplane and HBA/RAID hardware. Mixed protocols should be documented by bay number so field technicians do not replace a device with an incompatible type.

Two OCP ports can separate storage traffic from client or cluster traffic. They can also be bonded for resilience, though link aggregation does not protect against a shared switch failure unless links reach separate devices. The two ordinary PCIe slots may take storage controllers or additional network adapters.

Redundant 2000 W supplies do not mean the server consumes 2000 W continuously. They define power capacity. Facilities planning should use the ordered CPU, DIMM, drive and adapter load, then check the remaining capacity with one supply or feed unavailable.

Configuration Guidance

Processors: avoid selecting two maximum-core CPUs by habit. Storage services may need frequency, memory bandwidth or encryption performance more than raw cores. Virtualisation hosts can use high core counts when licences and VM density justify them.

Memory: populate both sockets evenly and cover all channels. Filesystems and databases often use RAM for caching and metadata. Leave capacity for recovery processes rather than sizing only for steady state.

Drives: choose enterprise media by endurance, latency consistency and power-loss protection. Keep firmware families controlled. Separate boot devices from the data set and plan how failed media will be identified and replaced.

Controller: NVMe can attach directly, while SAS needs an HBA or RAID card. Hardware RAID may be wrong for software-defined storage that expects direct device access. Decide at architecture stage.

Networking: calculate usable storage throughput after replication and protocol overhead. One 100 Gb/s link can become the limit for a full NVMe set; faster ports are useful only when switches and clients can receive the traffic. The GPUMachines storage server category provides larger-bay alternatives.

Operations: monitor drive wear, inlet temperature, fan speed, corrected memory errors and network drops. Keep spare drives and cables that match the deployed variant.

Recommended Configuration Paths

Scale-out flash storage node

Use twelve enterprise NVMe drives, balanced dual EPYC CPUs, channel-complete memory and two high-speed OCP adapters connected to separate switches. Let the distributed-storage software manage protection across several identical nodes.

Virtualisation host

Prioritise RAM capacity, redundant boot, a protected local NVMe pool and network redundancy. Choose CPUs according to licence cost and VM density. Spread management and tenant traffic across separate interfaces.

Database or analytics server

Select frequency and memory capacity for the engine, use high-endurance NVMe with predictable latency, and keep backups outside the node. Validate NUMA placement with the real query or ingest workload.

AI data-preparation node

Use enough CPU cores for decode and transform stages, fast local scratch, and network links sized to feed the downstream GPU systems. Monitor pipeline throughput from source storage to GPU consumption rather than testing the node alone.

Alternatives and Related Systems

The RS720A-E13-RS24U offers twice the front-bay count in 2U and more expansion room. It may be a better fit when storage density and service access matter more than rack units. Single-socket storage servers reduce power and NUMA complexity for smaller deployments.

For cluster planning, Designing Storage for GPU Clusters explains the roles of local NVMe, shared filesystems and object storage. The AI storage starvation guide covers throughput, metadata and checkpoint behaviour around GPU workloads.

Buying Through GPUMachines

GPUMachines can check the CPU, DIMM, drive, backplane, HBA or RAID, OCP card and PSU combination against the current ASUS options. The quote review should include expected usable capacity, network egress and the data-protection model, not only hardware quantities.

Deployment planning can cover rack depth, rails, power feeds, switch ports and cabling. Hosted or leased options may be considered where facility density is the limiting factor; availability and terms depend on the final specification.

FAQ

Is the RS700A-E13-RS12U a GPU server?

No. It is a dense dual-socket CPU and storage platform. Its expansion slots are intended for network, storage or specialist adapters rather than a bank of high-power GPUs.

Can all twelve bays use NVMe?

ASUS documents a 12-NVMe arrangement. Other configurations mix NVMe with SATA or SAS, and SAS requires an appropriate controller. Confirm the exact backplane map.

How much RAM can it use?

The platform has 24 DDR5 DIMM slots. Capacity and speed depend on the chosen EPYC CPUs, DIMM type and population. Populate channels evenly before chasing maximum capacity.

Is one node enough for software-defined storage?

No, not for node-level resilience. One server can run the software, but replication and failover require multiple failure domains.

Does it need two storage network cards?

Not always. Two OCP positions allow redundancy or separation, but the right design depends on traffic and switch topology. Avoid adding ports that share the same unprotected path.

Can it serve GPU training data?

It can participate in a storage cluster or preprocessing tier. One 1U node is unlikely to meet the resilience and concurrency needs of a large GPU estate by itself.

What should be checked before data-centre installation?

Confirm rack depth, rail clearance, redundant power feeds, inlet cooling, cable paths, switch ports and the service procedure for drives and fans.

Verdict

The ASUS RS700A-E13-RS12U is a high-density CPU and flash platform for teams that can use two EPYC 9005 processors and twelve fast drives in one rack unit. Its twin OCP positions make it easier to build a credible data path out of the chassis.

Choose it for density with a defined software and network plan. Choose 2U when bay count, acoustic behaviour or easier service matters more. And do not confuse local NVMe speed with a complete storage service.

Sources and Further Reading

Configure the ASUS RS700A-E13-RS12U with GPUMachines.

← Back to blog