.jpg)
A ten-GPU server changes the buying question. The issue is no longer whether enough cards fit; it is whether the host, storage, network and rack can keep all ten useful without turning one dense node into an expensive point of congestion.
The ASRock Rack 4U10G-GNR2/RF+ is built for that density. It provides ten full-height, full-length, dual-slot PCIe 5.0 x16 accelerator positions around a dual-socket Intel Xeon 6 host. ASRock Rack's current GPU qualification list names two exact ten-card configurations for the /RF+ suffix: NVIDIA H200 NVL and NVIDIA RTX PRO 6000 Blackwell Server Edition.
This is a source-backed technical review, not a hands-on benchmark. GPUMachines has not measured the server described here, so performance, acoustic and power claims stop at the manufacturer documentation and transparent engineering calculations.
There is also a naming trap. /RF+ is a separate model suffix, not punctuation added to the ordinary /RF name. The QVL keeps /RF and /RF+ on separate rows, while the current platform guide identifies the extended /RF+ cooling layout. Buyers should quote the complete suffix and check it on the order paperwork.
Browse current PCIe GPU servers. GPUMachines does not currently expose a public configurator for the /RF+ suffix, so an exact build should go through technical quotation rather than a link to the different /RF chassis.
The specification that matters
The headline figure is ten dual-slot GPUs, but the supporting platform explains whether that density is useful.
| Area | ASRock Rack 4U10G-GNR2/RF+ | |---|---| | Form factor | 4U rackmount, 867 x 438 x 176.5mm in the current platform guide | | CPU | Two Socket E2 (LGA4710) sockets for Intel Xeon 6700P, 6500P and 6700E processors | | CPU power boundary | Up to 350W per socket in the platform guide | | Memory | 32 DDR5 DIMM slots, 16 per CPU; RDIMM and MRDIMM | | Accelerator positions | Ten FHFL dual-slot PCIe 5.0 x16 slots through a PEX89104 switch | | Other expansion | One CPU-direct FHFL x16 slot, two front FHHL x16 slots and one OCP NIC 3.0 x16 position | | Front storage | 18 hot-swap 2.5-inch PCIe 5.0 x4 NVMe positions; up to 16 positions available for SATA/SAS with the required controller and backplane arrangement | | Internal storage | One PCIe 5.0 x4 M.2 position | | Base networking | Dual Intel i350 1GbE plus dedicated IPMI | | Power modules | 3+1 redundant 3200W 80 PLUS Titanium CRPS | | Cooling distinction | Five middle plus five rear hot-swap 80mm fans in the current /RF+ platform guide |
Three extra x16 expansion connectors appear alongside the ten accelerator positions. They do not make this a thirteen-GPU server. The ten dual-slot positions define the accelerator capacity; the remaining connectors serve devices such as a network adapter, storage controller or other single-slot card.
Exact GPU qualification, not family guesswork
ASRock Rack's support list records a maximum of ten NVIDIA H200 NVL cards for 4U10G-GNR2/RF+ and a separate maximum of ten NVIDIA RTX PRO 6000 Blackwell Server Edition cards. Both entries specify passive cooling and PCIe Gen5.
That is narrower than saying the chassis supports every Hopper card or every RTX PRO Blackwell model. H100 PCIe, H100 NVL, RTX PRO 5000 and smaller workstation cards may share family names or mechanical conventions, but they do not inherit an exact-suffix QVL row. A quotation should use the supported card identifier, the cooling style and the validated quantity together.
The two qualified paths serve different work patterns. H200 NVL provides 141GB of HBM3e per card and is aimed at memory-heavy AI and accelerated computing. RTX PRO 6000 Blackwell Server Edition provides 96GB of GDDR7 and can suit inference, rendering, visual computing and virtual workstation estates that need a server-qualified PCIe card. Software certification can matter as much as raw memory capacity, particularly for professional visualisation or virtualisation.
Neither option turns the server into HGX. The accelerator bank uses PCIe, so tightly coupled jobs do not receive an HGX baseboard's NVSwitch fabric. For many independent inference replicas, render tasks or research containers, that freedom is useful. For large training jobs that exchange data continuously across every GPU, compare the node with an HGX server before committing.
What ten 600W-capable positions imply
The product family advertises airflow intended for GPU and accelerator cards rated up to 600W. That statement describes the cooling design boundary, not guaranteed support for any unlisted 600W card.
A simple upper-envelope calculation shows why facility planning cannot wait until delivery:
- Ten cards at 600W create a 6,000W accelerator allowance.
- Two processors at the documented 350W platform limit add 700W.
- DIMMs, drives, NICs, the motherboard, PCIe switch and ten high-speed fans add further load.
- Three active 3200W modules in a 3+1 arrangement provide 9,600W of module capacity before redundancy, but PSU capacity does not equal wall draw.
This arithmetic is not a consumption forecast. GPU power limits, CPU choice and fan control can move the measured result materially. It does show that a fully populated node belongs in a properly engineered rack with the correct supply voltage, redundant circuits, PDU outlets, breaker headroom and heat-rejection plan. Factory acceptance should record idle, representative workload and worst-case observed power before the rack planner treats any figure as measured fact.
The 867mm chassis depth adds another practical constraint. Rear cable space, power connectors and bend radius must fit without blocking exhaust. A nominally deep rack can still fail the installation if the rear door, PDUs or cable managers consume the remaining clearance.
Xeon 6 selection and NUMA ownership
The host accepts two Intel Xeon 6 processors from the 6700P, 6500P and 6700E families. Core count alone is a poor way to choose between them. Tokenisation, retrieval, data decompression, simulation code, storage services and network processing create different host-side demands.
Two sockets also create NUMA boundaries. The PCIe switch and other expansion links have defined CPU ownership, so scheduler policy, process affinity and memory locality can affect data movement. A container that allocates host memory on the remote socket may deliver less consistent performance than the same workload placed with its accelerator and network path in mind.
Start with the number of concurrent jobs and their host work, then choose processors. An inference service with efficient GPU kernels may benefit more from balanced frequency and I/O than from the largest core count. A mixed CPU/GPU simulation could need substantial CPU throughput. Virtual workstations add management and per-user overhead. The finished design should be checked as a node, not assembled from independent maximum values.
Memory speed has population rules
The board has 32 DIMM slots, split 16 per processor. ASRock Rack's current platform guide lists RDIMM up to 6400 MT/s at one DIMM per channel and 5200 MT/s at two DIMMs per channel. MRDIMM reaches up to 8000 MT/s with one DIMM per channel, subject to a compatible processor and memory qualification.
That makes an 8800 MT/s MRDIMM option wrong for this exact platform. It also means a buyer cannot quote the maximum speed and maximum slot population in the same line without checking the population rule.
Capacity planning should include the operating system, containers, preprocessing, model cache, CPU-resident services and failure headroom. Populate both sockets symmetrically unless the intended topology gives a defensible reason not to. When memory bandwidth matters, preserve channels before jumping to the largest DIMM. For capacity-led workloads, a denser RDIMM arrangement may be sensible even at a lower data rate.
Eighteen front positions are not thirty-four bays
ASRock Rack describes 18 hot-swap 2.5-inch NVMe positions and up to 16 SATA/SAS-capable positions for the /RF+ layout when the required RAID or HBA hardware is fitted. The safe interpretation is a front layout with interface alternatives, not 18 NVMe bays plus another 16 physical bays hidden elsewhere.
An exact quotation must state which positions use NVMe and which use SATA or SAS, then enforce the common physical maximum. The controller and cable plan matters. NVMe devices attach through PCIe paths; SATA and SAS need the correct storage controller and supported backplane arrangement.
Active model cache, checkpoints and scratch data normally favour NVMe. SATA or SAS can suit capacity, retention or an established RAID design. The internal PCIe 5.0 x4 M.2 slot is better reserved for a boot or service image than treated as part of the main dataset tier.
Storage bandwidth also has to leave the node. Loading a model quickly from local NVMe solves one stage of the problem, but shared datasets, checkpoints and replica deployment can still bottleneck on the network. Measure the whole path.
Base networking is management-class
Dual Intel i350 1GbE ports and dedicated IPMI provide useful base connectivity. They are not the workload fabric for ten modern accelerators.
The OCP NIC 3.0 x16 position and spare PCIe slots give the architect choices. A standalone inference server may need high-speed Ethernet for client traffic and model distribution. A storage-heavy node may need a separate path to shared NVMe. Multi-node training can require low-latency Ethernet or InfiniBand plus switches, optics and a topology sized around the collective pattern.
No high-speed GPU fabric should be marked as included in the base server unless the exact bill of materials contains it. GPUMachines sizes adapters and switch ports after the workload, node count, storage location and fault model are known.
Workloads that fit the architecture
The /RF+ earns its rack space when a scheduler can feed many independent or loosely coupled workers. High-volume inference replicas are an obvious use because each card can serve a model instance or tenant. Rendering queues, simulation batches and research containers also divide cleanly when jobs do not need constant all-to-all GPU communication.
Virtual workstation pools can use the RTX PRO path, provided the virtualisation stack, display mode and application certification match the proposed card. The ASRock Rack RTX PRO 6000 Blackwell page distinguishes display-off compute use from display-on modes, which should be decided before deployment rather than left as an afterthought.
A multi-node private AI service can use each chassis as a dense worker, but that increases the blast radius of a node outage. Monitoring, spare capacity and maintenance scheduling matter more when ten accelerators disappear with one server.
Cases where a smaller server wins
Do not select ten slots for a four-card workload on the promise of later expansion unless the power and rack plan genuinely supports that future state. A four- or eight-GPU server costs less to power, is easier to cable and reduces the number of accelerators affected by one chassis fault.
Choose HGX when tightly coupled training needs the fixed NVIDIA baseboard and high-bandwidth GPU fabric. Choose a tower workstation when one or two local users need interactive access and the room can tolerate the acoustics. Consider hosted hardware when the organisation lacks the electrical and cooling infrastructure for a dense 4U node.
The /RF+ is not a general answer to every GPU workload. It is a specialised density choice.
Procurement checks for the /RF+ suffix
Before issuing a purchase order, ask the supplier to record these details against the final bill of materials:
- Exact chassis suffix 4U10G-GNR2/RF+, with no substitution to /RF or the base 4U10G-GNR2.
- GPU manufacturer part number, cooling type, firmware and QVL quantity.
- Processor pair, thermal design boundary and NUMA-aware workload plan.
- DIMM type, per-channel population and resulting supported data rate.
- Front-bay map showing the 18 physical positions and any SATA/SAS interface alternatives.
- RAID or HBA model, cables and backplane mode where SATA/SAS is required.
- OCP or PCIe network adapters, switch ports, optics and cables.
- Input voltage, redundant feeds, measured acceptance power and rack cooling allocation.
- Chassis depth, rear clearance and service route.
That list catches the expensive mistakes before the server reaches the loading bay.
Frequently asked questions
Does the ASRock Rack 4U10G-GNR2/RF+ support ten GPUs?
It has ten FHFL dual-slot PCIe 5.0 x16 accelerator positions. Current exact-suffix QVL rows list ten NVIDIA H200 NVL or ten NVIDIA RTX PRO 6000 Blackwell Server Edition cards. Other GPU models need their own exact qualification.
Why does the chassis have thirteen x16 slots but only ten GPUs?
Ten slots are full-height, full-length, dual-slot accelerator positions. One additional rear position is CPU-direct and single-slot, while two front positions are FHHL. Those connectors support other expansion devices and should not be counted as extra dual-slot GPU bays.
Are the /RF and /RF+ models interchangeable?
No. ASRock Rack lists them separately, and the current platform guide identifies a distinct /RF+ chassis and rear fan arrangement. Use the complete suffix in the quote, configuration and delivery check.
Can all 18 bays use NVMe?
The platform documentation lists 18 PCIe 5.0 x4 NVMe positions. It also describes up to 16 SATA/SAS-capable positions with the required controller arrangement. These are interface choices within the physical front layout, not 34 independent bays.
Does it ship with high-speed Ethernet or InfiniBand?
The base specification lists dual 1GbE and dedicated IPMI. High-speed workload or cluster networking is selected separately through OCP NIC 3.0 or PCIe expansion.
Is 8000 MT/s memory supported?
The current platform guide lists MRDIMM up to 8000 MT/s at one DIMM per channel, subject to processor and QVL support. It does not establish 8800 MT/s support.
Buying assessment
The ASRock Rack 4U10G-GNR2/RF+ has a defensible purpose: it puts ten qualified PCIe accelerators behind a modern dual-socket Xeon 6 host while retaining room for storage and high-speed networking. Buyers with parallel inference, visual computing or batch workloads can gain useful density from it.
The design is unforgiving of vague quotations. The exact suffix, GPU QVL, memory population, physical bay map and facility envelope need to be fixed before purchase. Get those right and the server can replace several lower-density nodes. Get them wrong and the tenth GPU becomes the least important problem.
Compare visible PCIe GPU servers, or review the distinct ASRock Rack 4U10G-GNR2/RF configurator if the six-bay SATA/SAS /RF model is the required variant. GPUMachines treats the two suffixes as separate products.