ASRock Rack 4U8G-GNR2/RF Review: Eight PCIe GPUs

The storage line decides whether an ASRock Rack 4U8G-GNR2/RF quote is accurate. ASRock offers four hot-swap Gen5 NVMe bays or 24 hot-swap SATA/SAS bays, not 24 drives of every interface. That single distinction changes the controller, drive list and role of the machine.

The GPU side is clearer. Eight full-height, full-length, dual-slot PCIe 5.0 x16 positions sit behind a dual-socket Intel Xeon 6 host. ASRock Rack's current support list names NVIDIA H200 NVL, RTX PRO 6000 Blackwell Server Edition and RTX PRO 4500 Blackwell Server Edition for the exact /RF suffix. It is a dense PCIe compute platform, not an HGX system with an eight-GPU NVSwitch fabric.

Configure the ASRock Rack 4U8G-GNR2/RF after deciding which factory storage build you need. The public selector currently represents the 24-bay SATA/SAS version; the four-bay NVMe version needs manual quotation until the catalogue can model alternative backplanes.

Exact model and evidence boundary

This technical review covers 4U8G-GNR2/RF. ASRock Rack also lists 4U8G-GNR2, 4U8G-GNR2/RF+ and 4U8G-GNR2/SUPER_4L. They share a platform family, but the suffixes are not interchangeable. Chassis depth, rear airflow hardware, drive arrangement and GPU qualification can differ.

The article uses ASRock Rack's current product page, its live GPU support list and its 2026 Q2 GPU server guide. It does not claim hands-on testing by GPUMachines. Final qualification still depends on the ordered revision, firmware and card list available at quotation.

| Area | Published position | What the buyer must decide | | --- | --- | --- | | Form factor | 4U rackmount, 867 mm deep | Rack depth, rails, rear service space and lift plan | | Processors | Two LGA 4710 sockets, Xeon 6500P/6700P/6700E | Core count, host services and balanced socket population | | CPU power | Up to 350 W per socket | Cooling margin and total rack load | | Memory | 32 DDR5 slots, 16 per CPU | Capacity, channel balance and RDIMM versus MRDIMM | | GPU positions | Eight FHFL dual-slot Gen5 x16 | Exact qualified card, quantity and power limit | | Extra expansion | One FHHL Gen5 x16 and one FHHL Gen5 x8 | NIC and storage-controller placement | | Storage option A | Four hot-swap 2.5-inch Gen5 NVMe | Fast local scratch or model cache | | Storage option B | 24 hot-swap 2.5-inch SATA/SAS | Capacity, HBA or RAID and protection policy | | Internal storage | One Gen5 x4 M.2 on CPU 0 | Boot or service volume | | Base network | Two Intel i350 1GbE plus dedicated IPMI | Add-in data fabric required for serious GPU work | | Power | Four 2700 W Titanium CRPS, 3+1 | High-voltage feeds, PDU connectors and failure margin |

Eight PCIe GPUs, not an eight-GPU NVSwitch domain

Each accelerator position receives a direct PCIe 5.0 x16 link. The layout accepts full-height, full-length, dual-slot passive server cards, and ASRock describes airflow for devices drawing as much as 600 W. For inference replicas, rendering frames, simulation batches or GPU virtualisation, that arrangement can be more useful than a fixed baseboard because operators can assign cards to separate jobs and replace one accelerator without treating the entire GPU assembly as one unit.

There is a hard boundary. PCIe provides the host connection; it does not give all eight cards the same high-bandwidth, all-to-all fabric found in an HGX H200, B200 or B300 platform. H200 NVL cards can use supported NVLink pairings, but a set of pairs still differs from an NVSwitch domain that treats eight GPUs as a tightly connected group.

That distinction should drive the purchase. Replicated inference often scales by placing one model worker on each GPU, then distributing requests across workers. Rendering and many engineering codes also split work into separate units. Both patterns fit this chassis. A training run that constantly exchanges gradients across all eight accelerators may lose more time to communication, so its owner should compare the complete application result against HGX rather than comparing GPU memory totals alone.

The current GPU qualification list

ASRock Rack currently lists three GPU families for the exact 4U8G-GNR2/RF:

• NVIDIA H200 NVL, up to eight passive 600 W cards, shown as NVIDIA-qualified.
• NVIDIA RTX PRO 6000 Blackwell Server Edition, up to eight passive 600 W cards, shown as NVIDIA-qualified.
• NVIDIA RTX PRO 4500 Blackwell Server Edition, up to eight passive 200 W cards, shown as ASRock Rack validated.

The status labels matter. ASRock describes NVIDIA qualification as thermal, mechanical, power and signal-integrity work for a named server and GPU combination. A lower-power card from the same brand cannot be assumed qualified simply because it fits. Firmware, auxiliary power, brackets and airflow can differ.

This is why a configurator should not expand RTX PRO Blackwell into every desktop or workstation card. Nor should it add H100 PCIe merely because H200 NVL appears in the same Hopper family. Keep the exact qualified parts visible; send any alternative to engineering review.

GPU quantity also affects the rest of the build. One or two cards leave large portions of an 867 mm, 4U, four-PSU chassis unused. A smaller server may consume less base power and be easier to place. Six to eight high-power cards make the case for this platform much stronger, provided the site can remove the heat.

Dual Xeon 6 host and the 32-slot memory plan

The two LGA 4710 sockets accept Intel Xeon 6500P, 6700P and 6700E processors. ASRock publishes a 350 W CPU ceiling for the family. P-core models can suit latency-sensitive host work, compilation and services that need strong per-core performance; E-core models trade per-core behaviour for higher core density and may fit highly parallel preprocessing or many concurrent services.

The correct host CPU is not automatically the largest. Eight GPUs performing inference may need tokenisation, retrieval, request handling, observability and storage work, but a model worker that spends most of its time on the accelerator can leave expensive host cores idle. Rendering schedulers and virtual workstation brokers create different loads. Profile the CPU-side pipeline before choosing two top-bin processors.

Both sockets should be installed for a full eight-GPU design because PCIe and memory resources divide across the platform. A one-socket experiment can leave slots or paths unavailable, and it creates uneven NUMA access even where the operating system can see a device. Record which CPU owns each GPU, service slot and memory bank before deploying a scheduler.

Thirty-two DIMM positions provide 16 slots per CPU. ASRock's current guide lists:

• RDIMM up to 128 GB per module, with up to 6400 MT/s at one DIMM per channel and 5200 MT/s at two DIMMs per channel.
• 3DS RDIMM up to 256 GB per module, under the same published channel-speed limits.
• MRDIMM up to 64 GB per module and 8000 MT/s at one DIMM per channel.

The 8000 MT/s ceiling is easy to miss because other LGA 4710 systems and parts catalogues include 8800 MT/s MRDIMMs. Socket compatibility does not qualify the faster module for this server. Until ASRock publishes an exact update, a product configurator should omit 8800 MT/s parts and quote a verified 8000 MT/s MRDIMM manually if required.

Host-memory capacity depends on the job. Model weights normally live in GPU memory during execution, but CPU RAM may hold input queues, preprocessed data, retrieval indexes, virtual machines, page cache and staging copies. Populate the two sockets symmetrically; then test memory bandwidth and NUMA placement with the actual framework.

The front backplane is a factory choice

ASRock's wording is precise: four hot-swap 2.5-inch NVMe bays or 24 hot-swap 2.5-inch SATA/SAS bays. It does not describe a 24-bay tri-mode front plane where every position can take NVMe, and it does not support adding four and 24 to produce 28 external bays.

The four-NVMe version suits a compute server that needs a small local scratch pool, fast model cache or checkpoint staging. Four Gen5 x4 drives can deliver far more local throughput than a 1GbE base network can carry. They still need a filesystem, endurance class and protection plan that fit the workload.

The 24-bay SATA/SAS version fits capacity-oriented local storage, VDI images, retained outputs or a large collection of moderate-throughput devices. ASRock requires an additional HBA or RAID card. The extra controller occupies one of the front service slots, draws power and needs the correct cables and firmware. Hardware RAID is one option; software-defined storage may prefer an HBA that exposes individual drives.

One internal M.2 2280/22110 position attaches to CPU 0 at Gen5 x4. It can carry the operating system, but one device does not provide boot redundancy. If the installation needs mirrored boot media, ask how the selected chassis revision and controller plan can supply it rather than treating the lone M.2 slot as a mirror.

The current GPUMachines schema cannot switch the whole front backplane as one product option. The least misleading public representation is the 24-bay SATA/SAS build with SATA and SAS sharing a combined maximum of 24. The four-bay NVMe version should remain a manual factory choice until the schema supports mutually exclusive drive layouts.

Two service slots have to carry the data path

Eight rear x16 positions belong to the GPUs. One front full-height, half-length Gen5 x16 slot and one front full-height, half-length Gen5 x8 slot remain for services. There are no published OCP NIC slots on this model.

On the 24-bay storage build, one service slot may be consumed by the HBA or RAID controller. That can leave a single slot for the workload network. The adapter, lane width and CPU attachment should therefore be decided before the server is ordered, not after the GPUs and drives have used every obvious path.

The motherboard includes two Intel i350 1GbE ports, shared between the front and rear physical access points, plus a dedicated IPMI port connected to the ASPEED AST2600 BMC. They are useful for provisioning, host management and low-rate traffic. They are not the production data plane for eight H200 NVL or RTX PRO 6000 cards.

Network sizing starts with the workload. Independent inference replicas serving small requests may need less east-west traffic than distributed training, but model deployment, telemetry and shared storage still create bursts. Rendering farms move scenes and frames. VDI adds display protocols and user storage. Select a qualified 100, 200 or 400GbE or InfiniBand adapter only after checking switch ports, optics, cables, transceivers, PCIe placement and operating-system support.

Two ports on one adapter do not remove every failure domain. If uptime matters, connect them to an appropriate pair of switches and test failover. Keep BMC access on a separate management network.

3+1 power needs a high-voltage rack

Four 2700 W Titanium CRPS modules operate as 3+1. At 200 to 240 V AC each module can deliver its full 2700 W rating; ASRock lists only 1000 W at 100 to 127 V. A low-voltage circuit therefore cannot support the same GPU build or the intended redundancy.

Eight 600 W cards account for 4.8 kW. Add two CPUs at up to 350 W each, 32 DIMM positions, drives, a high-speed NIC, a storage controller and the fan system. The rack plan must use the actual configured draw, redundancy policy and measured acceptance result. Four PSU labels do not mean 10.8 kW should be booked as normal load, and 3+1 does not help if all feeds terminate on the same failed PDU.

The server is 867 mm deep, 438 mm wide and 176.5 mm high. Cable connectors and bend radius add to the rear clearance. Fully populated, it will also be much heavier than the bare chassis. Confirm cabinet depth, rail compatibility, floor loading, lift method and service access.

Cooling deserves the same care. ASRock markets enhanced airflow for 600 W cards, but the data-centre supply temperature, pressure, blanking panels and neighbouring equipment decide whether that airflow reaches the server. A dense row needs a heat-rejection calculation in kW, not a count of rack units.

Workloads that fit the platform

The strongest cases share one trait: they can use PCIe accelerators as independent workers or supported pairs.

Inference operators can place a model replica on each card, reserve cards for different models or divide them among tenants. RTX PRO 6000 Blackwell's large frame buffer can suit visual computing, inference and selected fine-tuning workloads, while H200 NVL targets workloads that benefit from HBM3e capacity and bandwidth. Rendering farms can schedule frames or tiles per GPU. Research groups can isolate projects rather than asking every user to coordinate one fixed eight-GPU job.

It can also support GPU virtualisation where the chosen card, hypervisor and licence permit it. That use needs more than hardware: plan identity, quota, monitoring, image management and a way to prevent one tenant's storage or network traffic from degrading everyone else.

Do not buy this model for a one-GPU pilot unless expansion is genuinely funded and scheduled. A smaller 1U or 2U server removes base fans, power modules and unused risers. Do not choose it for an all-to-all eight-GPU training job without comparing HGX application performance. And do not place it in an office; the fan and electrical design belong in a data centre.

Acceptance checklist

Before approving a build, record the following:

1. Exact 4U8G-GNR2/RF suffix and current chassis revision. 2. GPU model, passive cooling, quantity, power limit and qualification status. 3. Two supported Xeon 6 processors and a socket-local GPU map. 4. Symmetric 32-slot memory plan, with no unqualified 8800 MT/s MRDIMM. 5. Four-bay NVMe or 24-bay SATA/SAS factory backplane. 6. HBA or RAID model, mode, cables and drive qualification where required. 7. M.2 boot role and any requirement for boot redundancy. 8. Data NIC, lane width, CPU attachment, switch port, optic and cable. 9. Separate BMC network and production access policy. 10. High-voltage feeds, connector type, PDU diversity and 3+1 failure case. 11. Rack depth, rails, rear clearance, installed mass and lift plan. 12. Sustained GPU, storage and network test with one supported PSU failure.

Acceptance should run the intended components at the same time. A GPU burn test alone can miss a NIC or storage thermal problem; a drive test with idle GPUs says little about the populated chassis. Log GPU temperature, throttling, corrected PCIe errors, CPU NUMA traffic, storage latency, network errors, fan response and power from the rack PDU.

FAQ

How many GPUs does the 4U8G-GNR2/RF support?

It has eight full-height, full-length, dual-slot PCIe 5.0 x16 GPU positions. The current ASRock support list names H200 NVL, RTX PRO 6000 Blackwell Server Edition and RTX PRO 4500 Blackwell Server Edition for the exact model.

Does it have 24 NVMe bays?

No. The factory storage choices are four hot-swap Gen5 NVMe bays or 24 hot-swap SATA/SAS bays. The 24-bay arrangement must not inherit NVMe from the four-bay option.

Can I combine the four NVMe bays with 24 SATA/SAS drives?

Not from the published standard layouts. The word or denotes alternative backplanes. A custom arrangement needs written confirmation from ASRock Rack before it appears in a quote.

Why are 8800 MT/s MRDIMMs excluded?

ASRock publishes MRDIMM up to 8000 MT/s for this platform. Parts listed for other LGA 4710 servers do not become qualified here automatically.

Is H200 NVL the same as HGX H200?

No. H200 NVL is a PCIe card that can use supported NVLink pairings. HGX H200 integrates eight SXM GPUs with NVSwitch, giving a different multi-GPU communication topology.

Is the onboard 1GbE enough?

It is enough for management and light traffic, but it is not a sensible production data path for eight current accelerators or 24 drives. Most deployments will need a faster PCIe network adapter.

What voltage does the server need?

Full 2700 W output per PSU requires 200 to 240 V AC. At 100 to 127 V ASRock lists 1000 W per module, which is insufficient for a dense eight-GPU configuration.

When should I buy a smaller server?

Choose a smaller model when the realistic requirement is one to four GPUs, rack power is constrained or expansion has no funded timetable. Empty slots still carry the cost and base power of a large chassis.

Verdict

The 4U8G-GNR2/RF is a strong eight-GPU PCIe server when the workload can divide across cards and the data centre can support several kilowatts of compute in 4U. Direct Gen5 x16 links, dual Xeon 6 sockets and qualified 600 W accelerator choices give it a credible role in dense inference, rendering and shared research infrastructure.

Accuracy depends on the details the old catalogue blurred. Storage is four NVMe or 24 SATA/SAS, not 53 positions. MRDIMM stops at a published 8000 MT/s. /RF, /RF+ and SUPER_4L are separate products. The production network must use one of only two service slots, and full PSU output needs a high-voltage rack.

Open the 4U8G-GNR2/RF configurator and submit the GPU model, storage-backplane choice, host-memory target, network fabric and rack power details for compatibility review.

Technical sources

• ASRock Rack 4U8G-GNR2/RF product page
• ASRock Rack current GPU support list
• ASRock Rack 2026 Q2 GPU server guide
• Intel Xeon 6 product information
• NVIDIA H200 product information
• NVIDIA RTX PRO 6000 Blackwell Server Edition

ASRock Rack can revise CPU, memory, GPU and component qualifications. Confirm the exact suffix, current support list and ordered backplane before purchase.