ASRock Rack 4U4G-GNR/HPR Technical Review

Four dual-slot accelerators or seven single-slot cards: that single word, or, decides what the ASRock Rack 4U4G-GNR/HPR can become. Add the two numbers together and the result is an imaginary eleven-GPU workstation. Read them as alternative mechanical layouts and the product makes sense: a short-depth 4U host for up to four high-power PCIe GPUs, or a denser set of narrower cards.

The 4U4G-GNR/HPR uses one Intel Socket E2 processor, eight DDR5 memory channels, seven principal PCIe 5.0 x16 positions and one additional PCIe 5.0 x8 position. ASRock Rack currently lists it for NVIDIA RTX PRO 6000 Blackwell Server Edition. It also includes onboard dual 10GbE, remote management, two Gen5 M.2 positions and a 2+1 redundant power system.

This is a source-based technical review, not a claim that GPUMachines ran benchmarks on a sample. The analysis checks the exact platform specification, identifies configuration dependencies and draws a clear boundary around what the current manufacturer material supports.

Configure the ASRock Rack 4U4G-GNR/HPR or compare other rack GPU workstations.

Verified 4U4G-GNR/HPR specification

| Area | Current source-backed specification | | --- | --- | | Product class | 4U rack GPU workstation | | Chassis depth | 710 mm with GPU fan assembly; 596 mm without it | | Processor | 1 x Socket E2, LGA 4710 | | Supported CPU families | Qualified Intel Xeon 6700P, 6500P and 6700E series | | CPU thermal limit | Up to 350 W, subject to the current support list | | Memory | 8 DDR5 DIMM slots, one DIMM per channel | | Memory types | RDIMM, RDIMM-3DS and MRDIMM | | Accelerator layout | Up to 4 FHFL dual-slot or 7 FHFL single-slot PCIe 5.0 x16 cards | | Extra expansion | 1 FHFL single-slot PCIe 5.0 x8 position | | Current named GPU qualification | NVIDIA RTX PRO 6000 Blackwell Server Edition | | M.2 storage | 2 x M-key 2280/22110, PCIe 5.0 x4 | | Fixed storage | Included cage for 2 x fixed 2.5-inch devices; final interface and cabling must be confirmed | | Onboard data network | 2 x Intel X710-AT2 10GbE RJ45 and 2 x Intel i210 1GbE RJ45 | | Management | Dedicated IPMI through ASPEED AST2600 BMC | | Power | Current product page: 2+1 redundant 2700 W 80 PLUS Titanium CRPS | | Cooling | 4 x 80 mm middle fans plus 2 x 80 mm rear GPU fans |

The exact suffix matters. Related ASRock Rack 4U workstation and GPU-server families can use another CPU socket, storage cage, network controller or PSU arrangement. A quote for 4U4G-GNR/HPR must follow documentation for that model rather than borrowing figures from a similar Turin, W890 or dual-socket chassis.

The slot map is alternative, not additive

Seven full-height, full-length PCIe 5.0 x16 positions give the chassis its two population modes. Four dual-slot accelerators consume adjacent space, leaving the mechanical arrangement shown by ASRock Rack. Narrow single-slot devices can occupy as many as seven positions. The additional x8 position serves a separate expansion need, but it does not turn the chassis into another dual-slot GPU location.

Width is only the visible constraint. Passive server GPUs need the chassis airflow to carry heat through their heatsinks, and a 600 W card also needs the right power leads, firmware support and operating conditions. Connector placement, neighbouring card clearance and cable routing can disqualify a layout that appears to fit by slot count.

The current ASRock Rack RTX PRO 6000 Blackwell server page identifies this exact platform and states support for four full-height, full-length dual-slot cards up to 600 W each. That is much stronger evidence than a generic statement that the chassis has PCIe slots. It links a named accelerator family to a named host platform.

The same source evidence was not found for H100, H200, B100, B200, GB200 NVL2 or GB300 NVL2 in this exact model. GPUMachines has removed those groups from the product record. That does not prove every other PCIe accelerator is impossible; it means the public evidence does not justify offering those cards as standard selectable parts. A special project can still proceed after ASRock Rack or the system integrator confirms the exact card, power mode, firmware and support terms in writing.

RTX PRO Blackwell suits independent GPU work

NVIDIA RTX PRO 6000 Blackwell Server Edition uses a dual-slot passive form factor, PCIe Gen5 x16 and 96 GB of GDDR7 memory. ASRock Rack specifies the card with a power envelope up to 600 W on its current qualification page. Four cards therefore provide four separately addressable accelerator resources with a large local memory allocation on each device.

That structure suits inference replicas, rendering queues, visualisation, virtual workstation pools, simulation and batch compute. Each job can stay on one GPU when its model or scene fits local memory, while schedulers can assign multiple independent tasks across the node. Some software can use peer communication between PCIe devices, but buyers should not equate this with an HGX NVSwitch fabric.

An HGX baseboard targets tightly coupled multi-GPU work through a purpose-built high-bandwidth fabric. The 4U4G-GNR/HPR targets PCIe expansion. For four independent inference services, the PCIe design may be the more economical and serviceable choice. For a training run that must behave like one large eight-GPU memory domain, an HGX system is the appropriate starting point.

This distinction also affects fault planning. An application spread across four PCIe cards may stop when any participating process fails, while independent replicas can tolerate one GPU being drained or serviced. The chassis does not decide that architecture; the orchestration and application do.

One Xeon 6 processor sets the host boundary

Socket E2, also known as LGA 4710, accepts qualified Intel Xeon 6700P, 6500P and 6700E processors in this platform. ASRock Rack sets a 350 W CPU limit. A processor sharing the socket or family name still needs current support-list approval and a compatible BIOS.

One socket can be an advantage. It avoids cross-socket NUMA traffic, reduces host component cost and leaves fewer thermal sources competing with the GPUs. Many inference, rendering and visualisation jobs need enough CPU to feed accelerators, prepare data and run services, but they do not scale with a second processor.

The trade-off is fixed. This workstation cannot later gain a second CPU, another eight or sixteen memory channels, or the memory capacity associated with a dual-socket server. CPU-heavy preprocessing, very large host-memory datasets and dense virtual-machine consolidation may justify another platform even when four GPUs are sufficient.

Processor choice should follow the host workload. High core count helps concurrent preprocessing, data transforms and many virtual machines; higher per-core performance may matter more for serial preparation steps or licence-bound engineering software. The right answer comes from the application profile, not the most expensive supported CPU.

The 350 W figure is a chassis ceiling, not a promise that every 350 W Xeon works with every GPU and ambient-temperature condition. Four accelerators can already impose a severe cooling load. Confirm the exact processor, fan assembly, firmware and inlet conditions on the final build.

Eight DIMMs mean eight-channel planning

The platform provides eight DDR5 DIMM slots at one DIMM per channel. It accepts qualified RDIMM, RDIMM-3DS and MRDIMM modules. ASRock Rack's current catalogue lists RDIMM and RDIMM-3DS speeds up to 6400 MT/s, while supported MRDIMMs can reach 8000 MT/s.

Those figures describe platform ceilings. Processor model, memory type, module construction, capacity, firmware and qualification determine the actual rate. A fast label on a DIMM does not force the system to operate at that number.

For bandwidth-sensitive GPU feeding, populate all eight channels with matched Micron enterprise modules. An eight-DIMM set keeps channel use even and makes future fault analysis easier. Using fewer modules can reduce initial cost, but every unpopulated channel gives up part of the available aggregate bandwidth.

Capacity planning needs to include more than model weights. Host memory can hold preprocessing queues, tokenisation data, render assets, virtual-machine state, page cache and orchestration services. A four-GPU workstation with too little host memory may leave expensive accelerators waiting while the operating system moves data.

MRDIMM can improve memory data rate on supported Xeon 6 configurations, yet it should not become an automatic default. The processor, board firmware, DIMM part number and population all need to match ASRock Rack's current qualification. Mixed memory types or opportunistic module additions invite lower speeds and difficult support cases.

Storage is small and deliberately local

Two M-key connectors accept 2280 or 22110 PCIe 5.0 x4 M.2 devices. They can serve as operating-system drives, local scratch space or a small cache tier. The chassis also includes a cage for two fixed 2.5-inch devices.

The sources reviewed do not state a universal interface and cable path for every fixed-cage build. GPUMachines currently retains a two-drive SATA entry because that was present in the product data, but marks it for bill-of-material confirmation. A valid quote must identify the drive interface, cable assembly and controller rather than treating two empty positions as proof of SATA, SAS or NVMe support.

The two M.2 devices are not U.3 hot-swap bays. They are separate motherboard-attached positions. The fixed 2.5-inch cage is another pool. Collapsing them into four interchangeable drives would misstate service access and electrical connectivity.

This limited local storage is reasonable for a workstation expected to use network storage. A replicated inference node might keep the operating system and a working model set locally, then pull artefacts from a shared repository. Rendering clients may cache active assets while completed frames return to central storage. Neither case turns the chassis into a twenty-four-bay data server.

Drive endurance deserves attention. Model caches and read-mostly assets differ from write-heavy scratch, checkpoint or render-spill workloads. Use enterprise media rated for the planned write volume, and decide whether two local drives need mirroring. A single fast boot drive remains a single failure point.

Onboard 10GbE is useful, but not a GPU fabric

Two Intel X710-AT2 10GbE RJ45 ports give the system a practical base data network. Two Intel i210 1GbE ports and dedicated IPMI management support control, provisioning and out-of-band administration. The ASPEED AST2600 BMC allows remote platform management without depending on the host operating system.

Ten-gigabit Ethernet is adequate for management, many virtual-desktop sessions and moderate asset movement. It can become the limiting resource when four GPUs repeatedly load large models or datasets from shared storage. At line rate, one 10GbE port carries no more than 1.25 GB/s before protocol overhead. Four accelerators can consume a large working set far faster than that link can refill it.

A faster NVIDIA Networking adapter can occupy an available PCIe position when the storage and switch estate justify 100, 200 or 400 GbE, or an InfiniBand fabric. This consumes slot space and adds transceiver, cable, switch and thermal requirements. Reserve the network path while designing the GPU layout, not after every visible slot has been assigned.

Network speed alone will not fix a slow data path. Shared storage throughput, metadata performance, CPU preparation, protocol tuning and application access pattern may sit below the link limit. Model the whole route from storage media to GPU memory.

Power arithmetic exposes the real constraint

The current ASRock Rack product page and Q2 2026 GPU server catalogue specify 2+1 redundant 2700 W 80 PLUS Titanium CRPS. An older Q1 2026 Intel Xeon 6 deployment document stated 3200 W. The newer product-specific material should guide the catalogue, while the conflict remains in the admin notes so procurement confirms the ordered PSU assembly.

Four accelerators at 600 W each account for up to 2400 W. Add a processor approaching 350 W and the arithmetic reaches 2750 W before memory, storage, network cards, fans, voltage-regulation losses and the motherboard. This does not mean the platform is invalid; accelerator power modes, redundancy behaviour, PSU input, environmental limits and approved configurations can change the permitted operating envelope. It does mean that a four-card maximum cannot be turned into a four-card quote without checking the power conditions.

The PSU label is delivery capacity, not measured wall consumption. In a 2+1 redundant design, supplies share or reserve load according to the platform policy. Their ratings should not be added and described as normal server draw.

GPUMachines currently uses a 520 W barebone allowance for motherboard, fans, BMC and chassis electronics before selectable processors, memory, drives, NICs and GPUs. ASRock Rack does not publish this as a chassis-only measurement. It remains an engineering estimate for early rack planning and must be labelled accordingly.

Final rack calculations should use the ordered configuration's idle and sustained-load readings. Include PSU efficiency, peak workload, fan response and the facility's usable feed. Four high-power workstations can create a substantial rack load even though each chassis occupies only 4U.

Cooling, depth and service access

ASRock Rack lists four 80 mm middle fans and two 80 mm rear GPU fans. The chassis depth is 710 mm with the GPU fan assembly and 596 mm without it. The shorter figure is not the deployed depth for a passively cooled four-GPU build that needs the rear fan section.

Rack selection must allow for rails, power connectors, network cables and bend radius behind the chassis. A nominal 800 mm cabinet may be awkward once rear clearance and airflow containment are considered. Check the rail specification and usable internal depth rather than comparing only the chassis number with the cabinet's external depth.

Passive accelerators depend on the server's front-to-rear airflow. Blank adjacent positions, wrong fan firmware, recirculated exhaust or an unsuitable inlet temperature can reduce available power or cause throttling. Data-centre deployment needs blanking panels and sensible aisle management, even though ASRock Rack describes the product as a workstation.

Service planning matters too. A fixed 2.5-inch cage is not equivalent to a front hot-swap array, and densely cabled GPU power leads can increase repair time. Record which parts can be changed without removing the chassis and which intervention requires a maintenance window.

Workloads that fit the platform

Inference is the clearest match when each model instance fits within one GPU's memory or when software can divide work across independent PCIe devices. Four cards can serve separate models, tenants or replica groups, while the host handles scheduling, preprocessing and network services.

Rendering and visual computing also fit the geometry. The workstation can assign cards to separate render jobs or use supported multi-GPU software, and RTX-class features align with professional graphics and simulation applications. Licensing, remote-display requirements and card display modes still need checking for each software stack.

Virtual workstation pools can use the platform where the chosen NVIDIA software, hypervisor and card licensing support the intended partitioning. Capacity planning should use measured user profiles. A CAD engineer, a video editor and a light office VDI session do not consume the same framebuffer, CPU time or network bandwidth.

Batch compute can benefit when jobs scale across independent GPUs or run one job per card. The short-depth chassis may help in smaller data rooms and edge-adjacent facilities, provided those sites can supply the electrical feed and cooling demanded by the selected accelerators.

Cases where another system is better

Do not choose the 4U4G-GNR/HPR for an eight-GPU training job that expects NVSwitch. It has neither an HGX baseboard nor an integrated GPU fabric. A four-GPU maximum also limits scale-up before the application has to cross the network.

A dual-socket PCIe server may be better when preprocessing, simulation or virtualisation needs more host cores and memory channels. A storage-heavy GPU server makes more sense when the node requires many hot-swap NVMe devices. Buyers who need a GPU not listed for this exact platform should obtain written qualification or select a server whose public QVL already covers it.

There is also no virtue in filling all four accelerator positions if one or two cards satisfy the workload. Fewer GPUs reduce purchase cost, rack power and thermal load, and they leave more margin for network expansion. Size the build from measured memory and throughput needs.

Procurement checklist

Before releasing an order for the exact 4U4G-GNR/HPR, confirm:

• The selected Xeon appears on ASRock Rack's current support list and remains within the 350 W chassis limit.
• Eight matched Micron enterprise DIMMs follow the board's population and qualification rules when full channel bandwidth is required.
• The GPU model, quantity, power mode and firmware have exact-platform approval; current public evidence names RTX PRO 6000 Blackwell Server Edition.
• The slot map accounts for card width, the separate x8 position and any high-speed network adapter.
• The two M.2 devices and two fixed 2.5-inch positions have the intended interface, cable path and service method.
• The ordered PSU model, input voltage and redundancy mode support the complete configuration despite the older 3200 W document conflict.
• Rack depth, rail fit, rear cable clearance, inlet temperature and cooling capacity have been checked for the GPU fan assembly.

That evidence should remain with the final bill of materials. A family name in a selector is not a substitute for a qualified configuration.

Frequently asked questions

Does the 4U4G-GNR/HPR support eleven accelerators?

No. ASRock Rack states up to four dual-slot or seven single-slot FHFL PCIe 5.0 x16 cards. These are alternative mechanical layouts. The server also has one single-slot PCIe 5.0 x8 position for another supported expansion card.

Which GPU does ASRock Rack currently name for this workstation?

The current ASRock Rack RTX PRO Blackwell server page lists NVIDIA RTX PRO 6000 Blackwell Server Edition for the exact 4U4G-GNR/HPR. Other GPU families require exact-platform evidence or written approval before quotation.

Is this an HGX server?

No. It is a PCIe GPU workstation. It does not provide an HGX baseboard or NVSwitch fabric, so it suits independent or PCIe-connected accelerator workloads rather than applications that require an integrated eight-GPU fabric.

How much memory can it use?

The chassis has eight DDR5 DIMM slots at one DIMM per channel. ASRock Rack's catalogue lists qualified RDIMM, RDIMM-3DS and MRDIMM types with capacity limits by module type. Confirm the selected Xeon, BIOS and exact Micron part numbers on the final build rather than deriving a total from unqualified modules.

Does it have hot-swap U.3 storage?

The reviewed sources specify two PCIe 5.0 x4 M.2 positions and an included cage for two fixed 2.5-inch drives. They do not establish a four-bay U.3 hot-swap pool. The fixed-cage interface and cabling need bill-of-material confirmation.

Is 2700 W the server's expected power draw?

No. It is the current PSU rating. Actual draw depends on the selected parts and workload. GPUMachines uses a separate 520 W barebone estimate for early configuration work, but delivered-system measurements should replace it in rack planning.

Buying assessment

The ASRock Rack 4U4G-GNR/HPR has a clear job: host up to four qualified dual-slot PCIe GPUs around one Xeon 6 processor in a relatively short 4U chassis. Current manufacturer material makes RTX PRO 6000 Blackwell Server Edition the defensible accelerator choice, while the eight-channel memory design and onboard 10GbE keep the base platform proportionate.

Its limits are equally plain. Four dual-slot and seven single-slot are alternative layouts. Local storage is modest, the fixed-drive interface needs confirmation, and this is not an HGX fabric. Power also deserves a written build check because four 600 W cards can approach the capacity boundary once host components are added, while older and current ASRock Rack documents disagree on PSU wattage.

For inference replicas, rendering, professional visualisation and batch workloads that fit independent PCIe GPUs, those constraints are manageable and the single-socket design can avoid needless host cost. Buyers needing NVSwitch, eight GPUs, more host memory channels or a broad hot-swap storage plane should start with another platform.

Configure the ASRock Rack 4U4G-GNR/HPR with a checked CPU, memory, GPU, storage and network plan. The rack GPU workstation category provides the nearest catalogue comparison.

Official sources

• ASRock Rack 4U4G-GNR/HPR product page
• ASRock Rack Q2 2026 GPU Server catalogue
• ASRock Rack NVIDIA RTX PRO 6000 Blackwell server list
• ASRock Rack Q1 2026 Intel Xeon 6 deployment guide, retained to document the older 3200 W PSU statement

Sources were checked on 22 September 2026. ASRock Rack can revise support lists and factory options; confirm the exact processor, memory, accelerator, storage cabling, PSU, firmware and operating conditions on the final quote.