Twenty-four NVMe bays can make a server look like a storage appliance, but the ASUS RS720A-E13-RS24U is better understood as a data-heavy compute node. Its two AMD EPYC 9005 sockets and 24 DDR5 DIMM slots put substantial processing and memory beside the flash. Eight PCIe 5.0 expansion positions and two OCP 3.0 slots then provide the paths out to clients, peer nodes or a GPU cluster.
That balance creates several credible configurations. It can host virtual machines on local flash, run database services close to their data, perform CPU-side preparation for AI training, or form one member of a distributed storage system. It is not automatically the right answer for each of those jobs. A useful specification starts by defining who owns the data, how it survives a server failure and how quickly it must leave the chassis.
Executive Summary
ASUS RS720A-E13-RS24U is a 2U dual-socket server for AMD EPYC 9005 processors. ASUS specifies support for processors up to 400 W per socket, 24 DDR5 RDIMM slots, 24 front all-flash NVMe bays, eight PCIe 5.0 expansion slots and two OCP 3.0 positions. The platform uses ASUS's DC-MHS architecture and offers redundant 3200 W or 2700 W power-supply options, depending on the ordered build.
The server suits organisations that need high core density, broad memory bandwidth and a large local NVMe tier in the same 2U node. Private-cloud teams, database operators, distributed-storage engineers and AI platform groups can all make sound use of it, provided the chosen software has a clear placement and protection model.
It matters because 24 drives can generate more traffic than an ordinary server network can carry. The RS720A supplies enough I/O positions to address that problem, but the buyer still has to assign lanes, CPU locality and redundant switch paths correctly. It is excessive for light file serving, a small virtual-machine estate or a workload whose data belongs on shared storage. It is also the wrong chassis for dense GPU training.
Configure the ASUS RS720A-E13-RS24U through GPUMachines.
Key Specifications
| Area | Verified platform detail | | --- | --- | | Form factor | 2U rack server | | CPU platform | Dual AMD EPYC 9005 processors, up to the platform's supported 400 W socket envelope | | CPU sockets | 2 | | GPU support | Not positioned as a dense GPU platform; expansion is configuration-dependent | | Memory | 24 DDR5 RDIMM slots, one slot per memory channel across two processors | | Storage | Up to 24 2.5-inch all-flash NVMe hot-swap bays | | PCIe expansion | Eight PCIe 5.0 expansion slots plus two OCP 3.0 positions | | Networking | Selected through OCP 3.0 or PCIe adapters; speed and redundancy depend on the build | | Power | Redundant 3200 W or 2700 W supply options listed by ASUS | | Management and security | Remote management, platform root-of-trust functions and optional TPM support, according to configuration | | Best-fit workloads | Virtualisation, databases, analytics, distributed flash storage, AI data preparation and private cloud |
The figures describe the platform ceiling. Backplane wiring, risers, processors, OCP cards and other adapters must be checked together because physical position counts do not prove that every maximum can be used at once.
Platform Highlights
- Two EPYC 9005 processors put many cores beside the data. That arrangement favours virtualisation, software-defined storage, analytics and preprocessing that can divide work across NUMA nodes. Software licensing and per-core behaviour may make a smaller CPU pair the better purchase.
- Twenty-four memory channels are available through 24 DIMM slots. A one-DIMM-per-channel layout keeps population planning relatively clear. Balanced populations matter; fitting plenty of capacity to one socket does not feed threads running on the other.
- The front panel is an all-NVMe data tier. Twenty-four hot-swap devices can service many queues at once. Endurance, power-loss protection and sustained latency deserve more attention than a clean sequential-read figure.
- Ten I/O positions give the network room to breathe. Eight conventional expansion slots and two OCP 3.0 positions can accommodate client, replication and management fabrics, subject to the ordered risers and CPU lane map.
- The 2U chassis provides more thermal and service space than a 1U node. It still needs disciplined rack airflow. Two high-power processors, 24 DIMMs, 24 drives and several fast adapters can keep the fan wall busy even without GPUs.
Our Technical View
The RS720A-E13-RS24U makes the most sense when compute and storage must scale together. A database can keep active tables and logs close to the processors. A virtualisation host can combine guest capacity with local flash. A storage daemon can use the CPUs for checksums, compression, encryption and protection work before traffic reaches the network.
That proximity is useful, though it concentrates consequences. Firmware maintenance takes both the application and its local media offline. A motherboard fault removes all 24 drives from service until the software routes around the node. If every server holds a very large share of a dataset, rebuild traffic after a failure can become more disruptive than the drive count suggests.
We would not specify this machine by starting with the largest EPYC CPUs and filling every bay. The sensible starting point is a throughput budget: application reads, writes, replication, rebuild traffic, backup and management. That budget determines the NIC count, switch ports and oversubscription ratio. CPU and memory choices follow the software's actual work.
The AMD platform gives buyers a wide range of core-count and frequency choices, but there is no single best SKU. Per-core licensed databases may favour fewer, faster cores. Distributed storage can benefit from more threads if checksum, compression or erasure coding consumes CPU. Virtualisation adds another constraint because failover capacity matters more than the density of one fully packed host.
For AI estates, this server belongs beside the GPU nodes rather than in place of them. It can ingest datasets, unpack archives, validate samples, tokenise text and stage checkpoints. A well-sized data service may improve GPU use more than adding another accelerator to a badly fed cluster.
Best-Fit Workloads
Distributed all-flash storage
Several identical RS720A nodes can provide a foundation for Ceph or another distributed storage system. Each drive should be exposed in the form expected by the chosen stack, while replication or erasure coding spans separate servers and preferably separate rack or power domains. The network must carry foreground traffic and recovery traffic without making client latency unpredictable.
This is not a one-node storage answer. A single machine with redundant power still has one motherboard, one firmware domain and one maintenance schedule.
Virtualisation and private cloud
Dual EPYC processors and 24 memory channels support dense virtual-machine estates. Local NVMe can hold active guests, caches or a hyperconverged data layer. NUMA-aware placement helps large guests avoid unnecessary traffic between sockets, while cluster admission control must reserve enough capacity to absorb a failed host.
Databases and data services
Transactional databases can use low-latency enterprise NVMe for logs, tables and indexes. Analytical engines may value core count and memory bandwidth. Both need testing under sustained mixed reads and writes, including compaction, checkpoint or recovery activity; a short test against empty drives says little about that behaviour.
AI data preparation
Training pipelines often spend CPU time decoding images, extracting documents, tokenising text and checking data quality. The RS720A can run that work close to a broad flash tier, then feed GPU nodes over high-speed Ethernet or InfiniBand. The design succeeds only when egress and shared-filesystem paths match the rate at which GPUs consume batches.
Search and indexing
Search platforms combine large memory maps, frequent reads and background write activity. Fast media and many CPU cores suit that mix. Drive endurance, compaction traffic and node replacement policy need to be designed before index growth exposes them.
Who Should Consider It
Consider the RS720A-E13-RS24U when a 2U node must combine dual-socket EPYC compute with a large all-flash front end. It fits data-centre teams that already automate firmware, monitoring and clustered failover, as well as organisations building repeatable storage or private-cloud nodes.
The best buyer knows the intended software stack and can answer four practical questions: how the drives are presented, which traffic uses each NIC, where protection occurs and what remains available while one complete server is down.
Who Should Not Buy It
A small business file server does not need this density. Fewer drives and one socket reduce acquisition cost, idle power and operational work. Bulk archives should usually sit on capacity media or object storage rather than 24 premium NVMe devices.
Do not choose it merely because a line of the specification mentions many PCIe positions. Multi-GPU training servers need qualified power cables, accelerator spacing, directed airflow and often NVLink or NVSwitch. This chassis is built around CPU, storage and general I/O.
Teams without a distributed-storage operator should compare a supported storage appliance or managed service. Open-source software can work very well, but somebody still owns upgrades, rebalance behaviour, drive qualification, telemetry and emergency recovery.
Architecture Notes
CPU and NUMA placement
Each EPYC processor owns its memory channels and part of the platform I/O. Threads that repeatedly reach memory or a NIC attached to the other socket pay an extra hop. Hypervisors, databases and storage services should keep memory allocation, interrupt handling and worker threads near the relevant devices where the software permits.
Maximum core count is not automatically maximum application speed. Frequency, cache behaviour, memory pressure and licence terms can all change the useful choice. A proof-of-concept should reproduce steady-state load and recovery rather than only measuring an idle node.
Memory population
Twenty-four DIMM slots map neatly to the memory channels of two EPYC sockets. Populate both processors symmetrically and use DIMMs from a supported set. A partial population can save money initially, but it may reduce channel bandwidth or make later upgrades awkward if the intended final layout was not planned.
Virtualisation sizing needs room for the hypervisor, filesystem caches and host failover. Storage software also uses RAM for metadata and buffers. Filling every gigabyte with guest allocations or application heaps leaves the system fragile under pressure.
NVMe layout and endurance
Drive capacity is only one variable. Write endurance, power-loss protection, firmware consistency, thermal limits and replacement stock all affect the service. Mix-and-match devices can show different latency during garbage collection, which makes a distributed node harder to predict.
Some workloads want direct drive access; others use software mirroring, a host filesystem or a storage controller. Decide that layer first. Stacking hardware RAID under a distributed system without a clear reason can hide drive health and complicate recovery.
Network design
Twenty-four NVMe drives can outrun several ordinary network ports. Calculate usable storage throughput after protection overhead, add replication and rebuild demand, then size client-facing and backend links. Dual paths should terminate on separate switches when service availability requires it.
Management traffic deserves its own logical or physical path. A congested storage fabric should not prevent an operator from reaching the baseboard controller. For AI clusters, keep the storage design coordinated with the GPU fabric instead of assuming that every packet belongs on one network.
Power, cooling and service
Redundant power supplies protect against a supply or feed failure only when the rack wiring and load permit one remaining unit to carry the server. Model the configured draw at full processor, drive and adapter load. Check the PDU phase plan and the site voltage required for the selected supply output.
Front-to-back airflow needs an unobstructed cold aisle. Cable management at the rear must leave adapters and supplies serviceable. A 2U chassis is easier to handle than many dense 1U systems, but a fully populated server is still heavy enough to require safe lifting practice and suitable rails.
Configuration Guidance
Choose processors from the workload backwards. Frequency-sensitive databases and commercial per-core licences can make medium core counts attractive. Virtualisation and software-defined storage may use more cores, but reserve capacity for host failure and background recovery.
Populate memory channels deliberately. Start with an equal DIMM count and capacity on each processor. Document the final target before buying the first set, because replacing undersized DIMMs later can cost more than selecting the right population at order time.
Separate boot from data. Use mirrored boot media appropriate to the server design, then reserve front NVMe for application or storage duties. Define namespaces, filesystems and protection rules consistently across nodes.
Select enterprise NVMe by duty cycle. Read-intensive media can suit immutable datasets or caches. Database logs, metadata and write-heavy distributed storage may require greater endurance. Keep firmware versions controlled and maintain tested spares.
Treat the NICs as part of the storage system. OCP 3.0 can simplify servicing, while PCIe slots permit extra fabrics or security devices. Confirm lane width, CPU ownership, port speed, optics and switch compatibility before fixing the riser order.
Plan for degraded operation. Test a failed NIC path, one power feed, a full server outage and drive replacement. Rebuild time and application latency during those events reveal more than an ideal-state throughput result.
Recommended Configuration Paths
Distributed flash-storage node
Use several identically populated nodes, balanced EPYC processors, one DIMM per channel where budget permits, enterprise NVMe with matched endurance and redundant high-speed network paths. Present drives according to the storage software's guidance, then spread protection across server and rack failure domains.
Virtualisation host
Choose CPUs around guest density and licence cost, fit enough RAM for normal load plus host-failure headroom, and use protected local or distributed NVMe. Separate tenant, storage, migration and management traffic logically, with physical separation where utilisation or policy requires it.
Database and analytics server
Prioritise predictable per-core behaviour, memory bandwidth and high-endurance drives. Place logs and data according to the database vendor's guidance, maintain backups on another system, and test checkpoint plus recovery behaviour under realistic concurrency.
AI data-preparation node
Select cores for decode and transformation work, enough RAM for active datasets and a striped flash tier for staging. Size network egress against the aggregate GPU demand. Permanent datasets and checkpoints should remain on a protected shared service rather than depending on one local node.
Alternatives and Related Systems
The Intel-based RS720-E12-RS24U offers a related 24-NVMe 2U design for estates tied to Xeon 6 qualification. A single-socket storage server may provide better cost and power efficiency when software does not benefit from two NUMA domains.
Browse the GPUMachines storage server category for other bay counts and processor layouts. Designing Storage for GPU Clusters explains the division between local scratch, shared filesystems and object storage, while the AI training storage guide examines checkpoint and concurrency pressure.
Buying Through GPUMachines
GPUMachines can review the selected EPYC processors, DIMM population, backplane, NVMe media, risers, OCP cards and network adapters against the current ASUS configuration. The resulting bill of materials should state lane use, storage protection, usable capacity and the expected behaviour after a drive, NIC or node failure.
Deployment planning can cover rack depth, rails, power feeds, switch ports, optics, cabling and firmware baselines. On-premise, hosted and leasing choices depend on the final build and current commercial terms; none should be assumed from the chassis specification alone.
FAQ
Does the RS720A-E13-RS24U support 24 NVMe drives?
ASUS lists 24 front all-flash NVMe bays. GPUMachines should confirm the precise backplane and cabling for the ordered configuration.
How much RAM should be fitted?
Capacity follows the application, but the population should remain balanced across both processors and their memory channels. Include host, cache and failover headroom rather than sizing only for the normal application footprint.
Is this a good Ceph server?
It can make a strong all-flash Ceph node when deployed as part of a multi-node design with suitable direct-drive presentation, network bandwidth and failure-domain rules. One server alone does not provide distributed storage availability.
Should storage and client traffic use separate networks?
Separation helps when replication, recovery or client traffic can saturate a shared link. The right design may use separate physical fabrics or controlled VLANs, depending on throughput, policy and switch capacity.
Can GPUs be installed?
Some expansion combinations may accept accelerators, but ASUS positions this model as a rack and storage server. Dense or high-power GPU work belongs in a server with qualified accelerator power, spacing and airflow.
Are 3200 W power supplies always required?
ASUS lists 3200 W and 2700 W redundant options. The correct units depend on processor, drive, memory and adapter choices, plus input voltage and redundancy targets.
Does 24 NVMe mean the server needs 400 Gb/s networking?
Not automatically. Estimate sustained application, replication and rebuild traffic first. Some deployments need several fast ports; others remain limited by software, protection overhead or client demand.
Can GPUMachines help with a clustered deployment?
Yes. GPUMachines can review node configuration, network interfaces, switch connectivity, rack power and the relationship between storage nodes and GPU servers. Filesystem tuning and operational ownership should also be agreed with the software team.
Verdict
ASUS RS720A-E13-RS24U earns its place when a buyer needs two current EPYC sockets and 24 NVMe devices to work as one data node. Its strongest attribute is balance: processor, memory, flash and expansion capacity occupy the same 2U chassis without turning it into a token GPU server.
That density must serve an architecture. If the network, protection policy or failure-domain plan remains vague, a smaller node or a supported storage appliance will be easier to run. With those decisions settled, the RS720A can become a credible building block for private cloud, data services and AI storage pipelines.