The front of the ASUS RS720-E12-RS24U is almost entirely drive bays. Behind those 24 NVMe positions sits a dual-socket Intel Xeon 6 platform with 32 DIMM slots and a large expansion budget, so this 2U server can be built as far more than a flash shelf. It can run the database, virtual machines or data service that owns the drives.
That combination is useful when compute and storage need to scale together. It can also create an expensive island if the network is undersized or the workload would be better served by separate compute and storage nodes. The buying decision should start with failure domains and data movement, then move to processor SKUs and drive capacities.
Executive Summary
ASUS RS720-E12-RS24U is a 2U dual-socket server for Intel Xeon 6 Scalable processors. ASUS lists 32 DDR5 DIMM slots, 24 front NVMe bays, ten PCIe expansion positions, two M.2 devices and ASMB12-iKVM management. The DC-MHS-based layout is intended to make CPU, I/O and service modules easier to standardise across data-centre deployments.
The platform fits virtualisation clusters, databases, analytics, high-density NVMe services and CPU-side data processing around AI infrastructure. Its 24-drive front end can provide local flash, a building block for distributed storage or a fast ingest tier. The expansion area leaves room for high-speed networking, storage control and other adapters.
It matters to buyers who prefer Intel software qualification or have standardised on Xeon 6. It is excessive for ordinary file serving, small VM estates or workloads that cannot exploit many memory channels and NVMe queues. Teams needing several GPUs should choose a chassis built around accelerator airflow and power rather than forcing cards into a storage-led design.
Configure the ASUS RS720-E12-RS24U through GPUMachines.
Key Specifications
| Area | Verified platform detail | | --- | --- | | Form factor | 2U rack server | | CPU platform | Intel Xeon 6 Scalable processors | | CPU sockets | 2 | | GPU support | Configuration-dependent expansion; not positioned as a dense GPU server | | Memory | 32 DDR5 RDIMM slots, up to 6400 MT/s at supported populations according to ASUS | | Storage | Up to 24 2.5-inch NVMe drives plus two internal M.2 positions | | PCIe expansion | Up to ten PCIe slots in the platform layout | | Networking | PCIe and OCP options, selected to suit the ordered configuration | | Power | Redundant power supplies selected for the final processor, drive and adapter load | | Cooling | Hot-swap fan-bar design for processors up to the supported thermal envelope | | Management | ASUS ASMB12-iKVM and dedicated management connectivity | | Best-fit workloads | Virtualisation, databases, analytics, all-flash storage nodes, CPU-side data preparation and private cloud |
Expansion counts describe the platform, not a guarantee that every slot, drive mode and adapter can be used together. Backplane, riser and network choices must be checked as one configuration.
Platform Highlights
- Thirty-two DIMM slots favour memory-heavy work. Dual Xeon 6 processors can be paired with a broad memory population for virtual machines, databases and analytics. Channel balance and supported DIMM speed matter as much as total capacity.
- Twenty-four front NVMe bays create a serious local data tier. The media can serve active databases, VM storage, caches or a distributed-storage daemon. Device endurance and protection policy decide whether it is production storage or fast temporary space.
- A 2U thermal envelope is easier to manage than 1U at similar CPU power. Larger heatsinks and a hot-swap fan bar give the system more cooling room, while the wider service space helps with cable and fan replacement.
- Ten expansion positions keep the platform adaptable. High-speed NICs, DPUs, storage adapters or specialist cards can be fitted according to the chosen risers. Physical count is only the start; lane width and CPU ownership need to match the use case.
- DC-MHS modularity supports repeatable servicing. Standardised host and management modules can reduce the variation between deployed nodes. Operations teams still need a controlled firmware baseline and spare-part list.
Our Technical View
RS720-E12-RS24U is best viewed as a data-intensive compute node. It places enough CPU and memory beside the drives to run demanding services locally, then gives those services room for high-speed network adapters. For databases and virtualisation, that can reduce the latency and cost of an external storage hop.
The same design can work as one node in a scale-out storage system. Twenty-four NVMe devices provide a large pool of queues and bandwidth, while two CPUs can handle checksums, encryption, replication and client protocols. A distributed filesystem may prefer direct access to each drive rather than hardware RAID. Architecture and controller mode must agree.
Intel standardisation is a genuine buying factor. Some organisations have certified software images, monitoring, security controls and licence agreements around Xeon. That operational fit may outweigh a small component-level advantage elsewhere. It should be stated directly rather than disguised as a universal performance claim.
The server is less convincing when a project combines every role. Running databases, VM storage, AI preprocessing and backup on one node can look efficient until maintenance or failure interrupts all of them together. Dense hardware does not erase failure domains.
Best-Fit Workloads
Virtualisation and private cloud
Thirty-two DIMM slots and two Xeon sockets suit VM density. Local NVMe can host latency-sensitive guests or cache a shared tier. Very large VMs should be aligned to NUMA boundaries where possible; a guest spanning both sockets may see different memory latency.
Transactional and analytical databases
Databases can use fast local media and large memory pools, especially for ingest, indexes and temporary work. Sustained latency under write pressure matters more than a short sequential test. Power-loss protection, drive endurance and backup design need equal attention.
Scale-out storage
The RS720 can act as a flash node in a distributed system. Replication or erasure coding should cross servers and, where possible, racks or power domains. One 24-drive node offers capacity, not node-level availability.
AI data ingest and preprocessing
CPU-side decode, tokenisation, augmentation and validation can occupy many cores and generate heavy storage traffic. The node can stage data before sending it to GPU servers. Measure the complete pipeline, because network egress can become the limit.
Search, indexing and content services
Search engines and content platforms often combine large memory maps, many small reads and background compaction. NVMe density and CPU capacity suit those patterns, provided write amplification and replacement rates are planned.
Who Should Consider It
The server suits data-centre teams already operating Intel infrastructure and needing a repeatable 2U node with many flash devices. It also fits organisations that want compute and storage in the same failure domain for performance reasons and have software that can replicate across nodes.
Buyers should have an explicit answer for how data survives a drive, controller, motherboard or complete server failure. Without that, 24 bays merely concentrate risk.
Who Should Not Buy It
Do not buy RS720-E12-RS24U for a small service that needs a few drives and modest RAM. A single-socket server will cost less, draw less power and avoid cross-socket tuning. Bulk archive data belongs on high-capacity HDD or object storage rather than expensive all-flash bays.
Dense GPU training needs another chassis. Expansion slots do not provide the GPU spacing, power harness and front-to-back airflow of a proper four- or eight-GPU server. Buyers who need only a shared storage target should also compare dedicated appliances and distributed designs before putting application compute on every storage node.
Architecture Notes
Dual Xeon processors split memory and PCIe devices between sockets. The operating system can schedule threads anywhere, but poor placement may send storage interrupts or application memory across the inter-socket link. Databases, hypervisors and storage daemons should be tested with NUMA policy enabled.
Thirty-two DIMM slots do not imply that a random population is sensible. Populate memory channels symmetrically and check the supported speed for the selected Xeon 6 processors and DIMM type. More channels can improve sustained data movement even when the application does not need the maximum capacity.
Twenty-four NVMe drives can exceed one network link. Approximate usable array throughput after software protection, then size OCP or PCIe network cards and switch ports around that figure. If clients connect through 25 Gb/s, a shelf of fast drives will spend much of its time waiting.
Storage protection depends on the software. Hardware RAID can simplify some deployments, while Ceph, ZFS and other systems may need direct device visibility or an HBA. Write cache settings and power-loss behaviour must be understood before production data arrives.
The 2U fan and PSU arrangement should be modelled at failure conditions. Losing one fan or supply changes thermal and electrical headroom. Rack power calculations must include both processors, 32 DIMMs, 24 drives and high-speed network adapters.
Configuration Guidance
CPU selection: choose Xeon 6 SKUs according to software certification, core licensing, memory need and per-core behaviour. Storage and database work may benefit from frequency and memory bandwidth more than the largest possible core count.
Memory: fill channels evenly across both sockets. Virtualisation capacity should include hypervisor overhead, caching and failover room. Database sizing should leave memory for the operating system and filesystem rather than assigning every gigabyte to the engine.
NVMe: use enterprise drives with power-loss protection and endurance appropriate to the write rate. Standardise firmware and spare capacity. Separate boot devices from application data and define what happens during a rebuild.
Networking: reserve enough slots and CPU-local lanes for data traffic. OCP 3.0 is useful for serviceability, while conventional PCIe can hold additional fabric or security adapters. Separate management from storage and tenant traffic.
Filesystem or RAID: decide whether protection happens in hardware, in the host filesystem or across several servers. Do not layer several parity schemes without understanding the write and recovery cost.
Monitoring: collect per-drive latency, wear, temperature and media errors, plus CPU locality, memory errors, fan health and network drops. Alert on tail latency, not only average throughput.
Recommended Configuration Paths
High-density virtualisation host
Use Xeon CPUs chosen around licence cost, a broad balanced memory population, protected boot and enterprise NVMe for active guests. Add redundant high-speed network cards and keep enough failover capacity elsewhere in the cluster.
Database server
Prioritise CPU frequency, RAM and low-latency high-endurance drives. Mirror or replicate according to the database design, and send backups to a different system. Test under sustained writes and recovery, not only clean reads.
Distributed flash-storage node
Deploy several identical servers with direct-access NVMe, high-speed dual-path networking and software protection across nodes. Keep control and client traffic separated where the storage stack benefits.
AI data-preparation node
Choose cores for decode and transforms, enough RAM for large working sets, a striped scratch tier and network bandwidth matched to the GPU cluster. Completed datasets and checkpoints should live on protected shared storage.
Alternatives and Related Systems
Single-socket 2U storage servers reduce power and may be easier to licence. The AMD RS720A-E13-RS24U offers a similar 24-bay idea around dual EPYC 9005 processors. The right choice follows software support and workload behaviour, not vendor allegiance.
Explore other drive counts in the GPUMachines storage server category. For a wider architecture decision, Designing Storage for GPU Clusters separates local scratch, shared filesystems and object storage. The AI training storage guide covers concurrency and checkpoint pressure.
Buying Through GPUMachines
GPUMachines can check the selected Xeon processors, DIMM population, backplane, NVMe media, risers, OCP cards and storage controller against the current ASUS configuration. The review should state usable capacity, protection method, network bandwidth and expected failure behaviour.
Rack planning can cover rail depth, power feeds, switch ports, optics and service clearance. Hosted and leasing routes depend on the final configuration and current commercial terms.
FAQ
Does the RS720-E12-RS24U support 24 NVMe drives?
ASUS specifies up to 24 front NVMe devices. The exact backplane and riser arrangement must be confirmed for the ordered server.
Is it suitable for a Ceph node?
It can be, provided drives are exposed in the way Ceph expects and several nodes provide replication or erasure coding. Network bandwidth and failure-domain design are central.
Why are there 32 DIMM slots?
The dual Intel platform supports a broad memory population for capacity and channel bandwidth. Actual speed depends on processor, DIMM type and population.
Can it run GPUs?
Selected expansion configurations may accept accelerators, but this model is storage-led. A dedicated GPU server is better for several high-power cards.
Does local NVMe remove the need for shared storage?
No. Local media can serve one host quickly. Shared storage, replication or backups are still needed when data must survive a node failure or be accessed by several machines.
What network speed should be selected?
Measure expected client and replication traffic. Twenty-four NVMe drives can justify 100 Gb/s or faster links, but the correct speed depends on protection overhead, concurrency and switch capacity.
Can GPUMachines review RAID and HBA choices?
Yes. The controller should be selected with the filesystem or distributed-storage software, because some stacks need direct drive access while others use hardware RAID.
Verdict
ASUS RS720-E12-RS24U is a capable 2U data node for organisations that need dual Xeon 6 compute, large memory capacity and 24 NVMe bays in one managed platform. It is strongest when software can use local flash and protect data across a wider cluster.
Do not pay for its density without a network and recovery design. A smaller single-socket host may be more efficient; a dedicated storage appliance may be easier to operate. The RS720 earns its rack space when compute and flash genuinely belong together.