.jpg)
The ASRock Rack 1U2N2G-AM5/2T is two servers in one enclosure, not one server with two GPUs. Each node has its own AM5 processor, four DIMM slots, one dual-slot GPU position, one M.2 drive, two 10GbE ports and dedicated management. The nodes share the 1U chassis, fan wall and redundant 1,600 W power system.
That architecture is useful when jobs are independent. Two inference replicas, render workers, development environments or hosted tenants can occupy one rack unit without sharing an operating system. It is the wrong design for a model that needs to pool memory or exchange tensors between two local GPUs.
Our verdict: buy it for node density and isolation, not for multi-GPU scale-up. Treat every CPU, memory and GPU figure as per node, then verify the chassis-level power and failure plan across both nodes.
This is a source-backed technical review. GPUMachines has not claimed laboratory benchmark results for every configuration. ASRock Rack support lists, processor limits, card fit, firmware and application behaviour must be checked for the exact build.
How the two-node chassis works
The 1U2N2G-AM5/2T contains two independent AM5 compute nodes. Each node boots its own operating system and exposes separate IPMI management. Software on node A does not see node B's RAM, M.2 drive or GPU as locally attached hardware.
A fully populated chassis therefore contains:
- two AM5 processors, one per node;
- eight DIMM positions in total, four visible to each node;
- up to two GPUs, one attached to each node;
- two M.2 drives, one per node;
- four 10GbE data ports and two dedicated management interfaces.
The GPUMachines configurator builds one node and multiplies that component set by the chosen node count. A selection of one CPU, four DIMMs and one GPU means that configuration in each populated node. It does not mean the two nodes share those parts.
This distinction also applies to capacity. Two nodes with 128 GB each provide 256 GB across the chassis, but no individual job can address 256 GB unless the application distributes itself over the network.
Why node density can be useful
Putting two servers into 1U can improve rack density for services which already scale horizontally. A load balancer can direct requests to either inference node. A render queue can hand frames to both. A hosting operator can assign one node to each tenant without running both inside one hypervisor.
The design may also reduce the amount of duplicated chassis metal and power hardware compared with two conventional 1U servers. That does not halve the operational work. Each node still needs firmware tracking, monitoring, operating-system maintenance, network configuration and application ownership.
There is a shared failure domain. The nodes have independent compute boards, but they occupy one enclosure and depend on the same chassis cooling and power subsystem. A service that must survive a complete chassis outage should spread replicas across at least two physical chassis, preferably across separate rack power and network paths.
CPU choice: AM5 with a 120 W chassis limit
ASRock Rack lists AMD EPYC 4005 and 4004 plus Ryzen 9000, 8000 and 7000 series support. The manufacturer's 2026 GPU server brochure gives this model a processor TDP limit of up to 120 W per node. That limit excludes some high-power AM5 parts even though they fit the socket.
EPYC 4004 and 4005 are the clearer server choices where 24-hour operation, platform validation, ECC memory behaviour and support lifecycle matter. AMD positions these families for entry-level servers and hosted services. Ryzen can still suit price-sensitive visualisation, build or compute nodes when the exact CPU and memory arrangement appear on the ASRock Rack support list.
Do not choose from core count alone. A single-GPU inference worker may need CPU time for tokenisation, decompression, API processing and network interrupts. Rendering can depend on scene preparation. A high-frequency processor may beat a larger core count for one application, while parallel data processing can prefer more cores.
Measure the host phase and stay inside the node's thermal limit. A 1U shared fan wall leaves less room for optimistic CPU assumptions than a large tower.
Four DIMM slots per node
Each node has four DDR5 UDIMM positions on the two-channel AM5 memory controller, allowing two DIMMs per channel. The public configurator now presents four slots per node rather than dividing them across the chassis.
Populate both channels symmetrically. Two larger DIMMs preserve two empty positions and may support a higher memory rate than four modules, depending on processor and DIMM organisation. Four modules can reach more capacity without buying the largest sizes, but the supported speed may fall at two DIMMs per channel.
This is a UDIMM platform. It does not use the high-capacity registered memory found in EPYC 9000 servers. A node needing very large host memory, twelve memory channels or multi-terabyte capacity has outgrown the purpose of this system.
Size memory for the largest job on one node. Include model staging, dataset cache, virtual machines, operating-system allowance and peak concurrency. Chassis total is a purchasing figure, not an addressable-memory figure.
One GPU per node, with a short qualification list
Each node provides one full-height dual-slot PCIe Gen5 x16 position. ASRock Rack also describes an alternative arrangement of two single-slot x8 devices per node. Mechanical fit, power connection, cooling direction and firmware support remain part of qualification.
The current public ASRock Rack GPU support list names two cards for this exact chassis: an AMD Radeon RX 7900 XTX Passive 24 GB and an NVIDIA RTX A2000, both listed with Ryzen 7000. That list is much narrower than the set of GPUs that physically use a dual-slot PCIe connector.
Do not infer support for H100, H200, B200 or a 600 W RTX PRO card from the slot name. A passive card also relies on the chassis airflow for cooling, while an active card brings its own fan behaviour. The exact part number, power cable and processor pairing need manufacturer approval.
The two node-local GPUs cannot form a direct NVLink or NVSwitch scale-up domain. Distributed software could send work over Ethernet, but that is a networked two-node job with separate memory spaces. It does not behave like two GPUs under one PCIe root complex.
Local storage and data placement
Each node has one internal M.2 position. There are no front hot-swap data bays. That keeps the design compact but restricts local capacity, redundancy and serviceability.
Use the M.2 device for the operating system, application and a replaceable local cache. Store durable models, datasets, checkpoints and tenant data on shared storage or another protected system. If a node needs mirrored boot drives or a large local NVMe tier, choose a platform built for those requirements.
Replacing the internal drive usually interrupts the node and requires chassis access. The maintenance plan should identify which node owns each drive and how to rebuild it from automation or backup.
Networking and management
Each node includes two 10GbE RJ45 ports driven by a Broadcom BCM57416 controller, plus dedicated IPMI. A full chassis presents four 10GbE data interfaces and two management interfaces to the rack.
Ten-gigabit Ethernet can serve client requests, shared storage and replication for many single-GPU tasks. It can also become the bottleneck when nodes repeatedly move large models, video or scientific datasets. Measure the transfer window and application traffic together.
The full-height expansion position is normally occupied by the GPU. Adding a faster NIC can conflict with that choice, while the alternative x8 slot arrangement changes card format and lane width. Buyers who require both a large GPU and a 100 or 200 Gb/s adapter should confirm a supported layout or use a chassis with dedicated network expansion.
Separate management from workload traffic. Label every port by node in the switch and asset records; the shared chassis name is not enough when diagnosing a failed link or rebooting one side.
Shared power and cooling
The chassis uses two 1,600 W Platinum common redundant power supplies and eight hot-swap 40 mm fans. Both nodes draw from that system. A resilient configuration must remain supported when one PSU or feed is unavailable according to ASRock Rack's power rules.
Calculate the total for two CPUs, two GPUs, eight DIMMs, two M.2 drives, both node boards, all network ports and the fan wall. Use device limits for the design check, then measure the assembled chassis under simultaneous node load. Testing one node at a time misses the shared thermal and electrical condition.
This is not an office server. Small high-speed fans create significant noise, and the chassis expects controlled front-to-rear airflow. Confirm rack depth for the 650 mm enclosure, rail fit, PDU connections and rear access before delivery.
Workloads that suit the design
Paired inference replicas
Each node can hold one model replica when the validated GPU has enough memory. A load balancer distributes requests, and one node can be serviced while the other remains available. Chassis-level resilience still requires another enclosure.
Rendering and visual computing
Independent frames, scenes or user sessions divide naturally by node. The GPU must appear on the support list and match the application driver and licensing requirements.
Hosted development nodes
Two teams or tenants can receive separate operating systems and hardware management. The isolation is stronger than placing both workloads in one host, although network, power and chassis ownership remain shared.
Edge data processing
The short-depth 1U format, AM5 CPUs and dual 10GbE can suit compact data-processing deployments where each stream runs on its own node. Environmental limits, dust, acoustics and service access still determine whether a conventional data-centre chassis fits the site.
When to choose another server
Choose a conventional two- or four-GPU server when one job needs several accelerators in the same operating-system image. Choose HGX when scale-up GPU communication controls performance. Choose a storage-rich platform when the workload needs multiple local NVMe drives, and choose an EPYC 9005 system when memory channels, PCIe lanes or CPU core count exceed AM5.
A pair of small workstations can also be a better answer when rack density is irrelevant and local acoustics, display output or frequent card access matter more. Density has value only when the rack, cooling and operating model use it.
Configuration profiles
Server-oriented inference pair
Use two supported EPYC 4004 or 4005 processors within the 120 W limit, balanced ECC UDIMMs, one qualified GPU and one enterprise M.2 drive per node. Keep model artefacts on shared storage and use the two 10GbE ports for service and storage traffic according to measurement.
Visual computing pair
Choose an ASRock Rack-validated graphics card, size memory for the largest scene on each node and confirm the remote graphics licence. Test both nodes under sustained graphics load to validate shared cooling.
CPU-first dense compute
The nodes can run without GPUs when the immediate requirement is high-frequency AM5 compute with IPMI and 10GbE. Compare the cost with a CPU-only multi-node chassis before reserving unused GPU positions indefinitely.
The ASRock Rack 1U2N2G-AM5/2T configurator now applies four DIMM slots per node. Component availability in the interface still needs to be checked against the current ASRock Rack CPU and GPU qualification lists before quotation. Compare the PCIe GPU server range if the project needs several GPUs in one node or more local storage.
Acceptance testing
Confirm the chassis revision, BIOS and BMC versions for both nodes. Record node serials, IPMI addresses, switch ports and M.2 ownership. Verify that the selected CPU remains within the 120 W limit and that the exact GPU part number is supported.
Run both nodes simultaneously with the production workload. Capture per-node CPU, memory, GPU, network and storage metrics plus chassis fan speed, temperatures and wall power. Then test a node reboot, loss of one data link and loss of one PSU feed according to the agreed resilience design.
If the service claims high availability, remove a complete chassis from the load balancer. That test exposes whether replicas were accidentally placed inside one shared failure domain.
Final assessment
The 1U2N2G-AM5/2T is a precise tool: two independent, modest single-GPU nodes in a dense enclosure. It can lower rack consumption for workloads that already scale by replica or job, while IPMI and dual 10GbE make each node operate like a small server rather than a desktop board.
Its constraints are equally precise. Memory, storage and GPUs remain node-local; CPU power stops at 120 W per node; local storage is one M.2 device; and the published GPU list is narrow. A valid quotation needs to respect those limits instead of filling every PCIe-shaped option in a generic catalogue.
Configure the ASRock Rack 1U2N2G-AM5/2T, then ask GPUMachines to verify both per-node component sets, the manufacturer support lists, shared power budget, firmware and rack deployment before ordering.
Sources
- ASRock Rack 1U2N2G-AM5/2T product page
- ASRock Rack 2026 Q2 GPU server guide
- ASRock Rack GPU support list
- AMD EPYC 4004 Series
- AMD EPYC 4005 and 4004 server overview
Sources checked 22 September 2026. ASRock Rack may revise CPU, GPU, memory and firmware support; confirm the exact system revision and qualified components before purchase.