ASRock Rack 4U8G-TURIN2 Review: 8-GPU EPYC Server

The ASRock Rack 4U8G-TURIN2 is a 4U dual-socket AMD EPYC server with eight double-width PCIe Gen5 GPU positions, 24 DDR5 DIMM slots and 24 front drive bays. Unlike many accelerator chassis, it combines dense GPU capacity with a substantial local SATA tier and four direct Gen5 NVMe bays.

Its name needs careful reading. 4U8G-TURIN2 describes the base model, while ASRock Rack sells related RF, RF+ and SUPER_4L variants with different cooling, networking or storage arrangements. A GPU qualified for an RF+ chassis should not automatically be treated as qualified for the base 4U8G-TURIN2. The exact suffix and current support list belong on the purchase order.

Configure the ASRock Rack 4U8G-TURIN2 on GPUMachines after checking the intended GPU against the exact-model QVL. Slot count alone does not confirm cooling, auxiliary power or firmware support.

Technical review summary

ASRock Rack supports two AMD EPYC 9005 or 9004 processors, including selected 3D V-Cache and EPYC 97x4 models, on socket SP5. The published CPU TDP limit is 500 W. Twenty-four DIMM sockets populate all twelve memory channels per CPU at one DIMM per channel, giving this platform a better host-memory layout than reduced-slot GPU servers.

Eight rear full-height, full-length, dual-slot positions provide PCIe Gen5 x16 links. A further front full-height, half-length Gen5 x8 position is available for an auxiliary adapter. There is also one OCP NIC 3.0 slot with a Gen5 x8 link. This lets the system retain a network path without necessarily consuming a GPU position, although the chosen OCP card still needs qualification.

The front storage layout is specific: four hot-swap 2.5-inch Gen5 x4 NVMe bays and twenty hot-swap 2.5-inch SATA bays. These are separate capability pools, not 24 universal NVMe bays. Two internal M.2 sockets support either Gen3 x4 NVMe or SATA, depending on the device and system configuration.

Onboard networking is limited to two 1GbE RJ45 ports through Intel i350, plus dedicated management through IPMI. A serious eight-GPU deployment will usually need a faster OCP or PCIe network adapter for storage, client traffic or multi-node communication.

Power comes from a 3+1 arrangement of 2,700 W 80 PLUS Titanium CRPS units. That does not mean every eight-card GPU set is supported. ASRock Rack lists higher-airflow RF+ and other related versions separately, and the current public GPU support list should be checked against the exact base model before any high-power accelerator is selected.

Verified system specification

| Area | ASRock Rack 4U8G-TURIN2 specification | | --- | --- | | Form factor | 4U rackmount | | Dimensions | 786 x 438 x 176.5 mm | | Processors | 2 x AMD EPYC 9005 or 9004, socket SP5 | | CPU power support | Up to 500 W TDP, subject to QVL and operating conditions | | Memory | 24 x DDR5 DIMM, twelve per CPU, 1DPC | | Memory types | RDIMM and RDIMM-3DS | | GPU expansion | 8 x FHFL dual-slot PCIe Gen5 x16 | | Auxiliary expansion | 1 x FHHL PCIe Gen5 x8 | | NVMe storage | 4 x hot-swap 2.5-inch Gen5 x4 bays | | SATA storage | 20 x hot-swap 2.5-inch SATA bays | | Internal storage | 2 x M.2, Gen3 x4 NVMe or SATA 6Gb/s | | OCP networking | 1 x OCP NIC 3.0, PCIe Gen5 x8 | | Onboard network | 2 x 1GbE RJ45 through Intel i350 | | Management | Dedicated IPMI and BMC management | | Power supplies | 3+1 x 2,700 W 80 PLUS Titanium CRPS |

The table describes the base 4U8G-TURIN2. Do not import specifications from 4U8G-TURIN2/RF, /RF+ or /SUPER_4L without confirming that they apply to the ordered chassis.

Why exact variant control matters

Manufacturer families often share a motherboard and exterior design while changing fans, drive backplanes, network rails or power accessories. Those differences are particularly important at eight-GPU density.

ASRock Rack describes the 4U8G-TURIN2/RF+ as a higher-airflow model for 600 W GPUs and accelerator cards. The base 4U8G-TURIN2 page does not make that same statement. The SUPER_4L variant has a different front storage arrangement, and RF models may include specialised network fabrics.

This creates a simple procurement rule: match the model string character for character across the quote, product page, GPU QVL, manual and delivered asset label. A configurator option is an invitation to check compatibility, not evidence that every variant supports it.

The current ASRock Rack GPU support list includes an exact base-model entry for eight NVIDIA RTX PRO 4500 Blackwell Server Edition cards. Other high-power cards appear against related TURIN2 variants. Support lists change, so the final decision should use the entry available at order time.

This does not mean the base chassis can never support another GPU. It means the integrator should obtain current confirmation rather than inferring support from slot size or from a neighbouring SKU.

Eight PCIe GPUs and direct host links

Eight Gen5 x16 positions make this a high-density PCIe platform. The accelerators remain add-in cards connected to the host PCIe fabric. There is no fixed HGX baseboard or NVLink Switch domain.

That architecture works well for:

• independent inference replicas and model-serving tenants;
• rendering and media-processing queues;
• virtual workstation pools;
• scientific or engineering jobs that scale across standard PCIe devices;
• fine-tuning jobs whose communication pattern fits PCIe or supported card-level links; and
• mixed CPU, storage and GPU pipelines that benefit from a large local drive tier.

It is less convincing for one job that constantly exchanges data across all eight GPUs. An HGX server provides a different scale-up architecture through SXM modules, NVLink and NVLink Switch. Count is not topology: eight PCIe cards and eight HGX GPUs solve different problems.

The direct CPU-to-GPU layout also creates two NUMA domains in a dual-socket system. GPU, NIC and storage locality can influence throughput. Map the delivered topology with operating-system tools and bind services where the application benefits.

CPU choice for an eight-GPU host

Two SP5 sockets provide enough cores, memory channels and PCIe resources for a wide range of hosts. EPYC 9005 offers Zen 5 and Zen 5c choices, while EPYC 9004 remains supported for buyers standardising on that generation.

CPU selection should follow the host workload:

• higher-frequency parts can suit serial preprocessing, orchestration and applications with limited thread scaling;
• higher core counts help with many independent services, heavy data preparation, compilation or virtualisation;
• 3D V-Cache models may benefit selected simulation and analytics codes; and
• dense-core processors can run substantial CPU work, but they also consume power and licence budget that may be better assigned elsewhere.

The published 500 W limit is a platform maximum, not a recommendation to install two top-power CPUs beside eight accelerators. Model the electrical and cooling envelope of the final combination.

Dual-socket systems require NUMA-aware memory and process placement. If a process on CPU 0 repeatedly feeds a GPU connected to CPU 1, traffic crosses the socket link. Schedulers and containers should preserve locality rather than viewing sixteen CPU and GPU endpoints as one flat pool.

Twenty-four DIMMs use the full EPYC channel count

Twelve DIMM slots per CPU allow one module on each SP5 memory channel. This matters for data preparation, CPU-based feature work, simulation, caching and any application that repeatedly moves data between host memory and eight accelerators.

Capacity should be derived from measurement. GPU VRAM does not replace system RAM. The host may need room for datasets, model copies, tokenisation, queues, virtual machines, file cache, checkpoints, monitoring and failover.

Populate both sockets symmetrically with modules from the current memory QVL. The supported data rate depends on processor generation, DIMM type, rank and firmware. A higher headline MT/s figure is useful only if the selected module and population can run it.

RDIMM-3DS can provide high capacity, but cost and latency characteristics may differ from ordinary RDIMM. Buyers should decide whether they need maximum capacity, maximum channel bandwidth or a balanced price per gigabyte.

Storage: four NVMe bays plus twenty SATA bays

The front panel is one of the machine's most useful features, but it is also easy to misconfigure. ASRock Rack publishes four hot-swap 2.5-inch NVMe bays with Gen5 x4 links and twenty hot-swap 2.5-inch SATA bays.

The correct configurator logic is therefore:

• a maximum of four front NVMe devices;
• a separate maximum of twenty front SATA devices; and
• two internal M.2 devices using either NVMe or SATA as supported.

The twenty SATA positions are not additional NVMe bays. They can hold enterprise SATA SSDs for capacity, model repositories, render assets, logs or a warm data tier. Four direct NVMe devices can provide a faster scratch, cache or checkpoint tier.

This split can be more useful than a small all-NVMe layout. Many GPU jobs need a fast active tier and a larger capacity tier in the same node. It is still necessary to model sustained throughput. Twenty SATA SSDs behind a controller or backplane do not deliver twenty independent full-speed paths to every GPU.

RAID, HBA and filesystem choices depend on the ordered backplane and protection requirement. The current exact product page calls the twenty bays SATA. Some broader family material refers to SATA/SAS layouts, so SAS should only be offered when the exact revision and required controller are confirmed.

Avoid a shared-pool configurator error

A common catalogue mistake is to expose all twenty front positions as both NVMe and SATA, then rely on a shared counter. That remains incorrect for this server because the hardware has four NVMe-capable positions and twenty SATA positions. The two pools need separate caps even though total chassis bays equal 24.

This distinction affects price as well as compatibility. An apparent twenty-drive Gen5 NVMe configuration would materially understate backplane, lane and controller requirements because the base chassis does not provide it.

Internal M.2 devices

Two M.2 sockets support Gen3 x4 NVMe or SATA 6Gb/s. They are suited to boot, hypervisor, recovery or service duties. Mirroring can improve operating-system resilience where the firmware and software stack support the chosen method.

Use enterprise devices with suitable endurance and power-loss protection. Internal M.2 media is less convenient to replace than front hot-swap drives, so avoid treating it as disposable high-write scratch unless the service plan accepts chassis access.

The interface is Gen3, not Gen5. That is adequate for boot and management but should not be confused with the four front Gen5 NVMe bays.

Network planning

Two onboard 1GbE ports are useful for installation, service traffic and lower-rate access. They are not a sensible primary data path for an eight-GPU system.

The OCP NIC 3.0 Gen5 x8 slot is valuable because it provides a dedicated network position without using one of the eight dual-slot GPU bays. Depending on the qualified card, this can carry storage, client or cluster traffic. The front Gen5 x8 position offers another adapter path.

Network speed should follow measured traffic. A single-node inference host may need fast client and model-loading access but little east-west GPU communication. Distributed training can require multiple high-speed rails and strict topology. Rendering may be limited by asset delivery and result writes.

Plan:

• BMC management on a separate administrative network;
• storage bandwidth for the four NVMe and twenty SATA tiers;
• client or application traffic;
• multi-node GPU communication, if required;
• switch ports, optics or cables and breakout mode; and
• NUMA locality between NICs and the GPUs they serve.

An OCP slot rated Gen5 x8 can carry a fast adapter, but the complete card, firmware and thermal combination still needs to appear on the relevant support list.

Power and cooling

The chassis uses four 2,700 W Titanium CRPS units in a 3+1 arrangement. Nameplate capacity is not measured wall draw, and redundancy should not be derived by simply multiplying ratings. The exact input voltage, PSU sharing policy and qualified component set govern available power.

Eight accelerators, two CPUs rated up to 500 W, 24 DIMMs, 24 front drives and high-speed networking can create a substantial rack load. Use the ordered configuration and measured worst-case workload in the rack power plan.

The base model and RF+ cooling statements differ. ASRock Rack explicitly describes RF+ as intended for 600 W GPUs. For the base model, check the current GPU QVL and do not assume that any 600 W passive card is approved merely because it is double-width.

Facility review should include:

• inlet temperature and any altitude derating;
• rack PDU voltage, outlet type and A/B capacity;
• sustained rather than idle power;
• front-to-rear airflow and blanking;
• cable clearance around 786 mm chassis depth;
• service access for four PSUs and 24 drives; and
• behaviour after a PSU or fan failure.

A full eight-GPU node belongs in a controlled data-centre environment. Fan noise and heat make ordinary office operation unrealistic.

4U8G-TURIN2 versus GIGABYTE G494-ZB0-AAP2

Both are 4U, dual-EPYC, eight-GPU PCIe servers with 24 DIMM slots, but their storage philosophies differ sharply.

GIGABYTE G494-ZB0-AAP2 devotes the chassis to accelerators and auxiliary x16 slots. It has only two internal M.2 devices and publishes qualified uniform eight-card sets including H200 NVL, RTX PRO 6000 Blackwell and MI350P under stated conditions.

ASRock Rack 4U8G-TURIN2 provides four front Gen5 NVMe bays, twenty front SATA bays, an OCP slot and a front x8 position. It is attractive when local capacity belongs in the same node. Its GPU choice must be tied to the exact base-model QVL rather than borrowed from RF variants.

Choose between them by qualified accelerator, storage pattern, network adapter placement, rack depth and service model. The broader chassis specification does not override the GPU support list.

Who should buy the 4U8G-TURIN2?

It is a good candidate for:

• inference fleets using eight qualified PCIe GPUs;
• render or media nodes that benefit from a large local SATA capacity tier;
• simulation and analytics pipelines combining GPUs with local datasets;
• private-cloud or virtual-workstation hosts;
• teams needing all 24 EPYC memory channels populated; and
• sites that want OCP networking without sacrificing a GPU bay.

When not to buy it

Do not choose it when the application requires one eight-GPU NVLink Switch domain, when one or two GPUs meet the need, or when the intended accelerator is qualified only for a different TURIN2 suffix.

It is also a poor fit if all 24 front bays must be NVMe. Only four are published as Gen5 NVMe on the base model. Buyers needing a larger all-flash NVMe tier should select a chassis designed for that backplane rather than forcing the catalogue to pretend these bays are universal.

Configuration checklist

Before approving an order, record:

1. the full chassis model including any suffix; 2. the exact GPU model, power class, quantity and QVL entry; 3. both CPU OPNs, TDP settings and firmware support; 4. a symmetric 24-DIMM population from the memory QVL; 5. no more than four front NVMe and twenty front SATA drives; 6. the two M.2 devices and their chosen NVMe or SATA mode; 7. any RAID or HBA requirement for the front storage tier; 8. the OCP and auxiliary PCIe network adapters; 9. switch, optics or cables and network separation; 10. rack power, cooling, depth and service access; and 11. operating system, accelerator driver, runtime and acceptance tests.

The acceptance plan should exercise sustained GPU load, CPU-to-GPU transfer, local storage, network throughput, thermal behaviour and hardware error reporting. A short boot test cannot validate an eight-GPU production node.

Frequently asked questions

How many GPUs fit in the ASRock Rack 4U8G-TURIN2?

The chassis has eight full-height, full-length, dual-slot PCIe Gen5 x16 GPU positions. The exact accelerator and power class must be approved against the current base-model GPU QVL.

Does it support AMD EPYC 9005?

Yes. It supports two EPYC 9005 or 9004 processors on SP5, including selected 3D V-Cache and EPYC 97x4 models.

How many DIMM slots are available?

Twenty-four, with twelve slots per CPU at one DIMM per channel. It supports qualified DDR5 RDIMM and RDIMM-3DS modules.

Are all 24 front bays NVMe?

No. Four bays are hot-swap 2.5-inch Gen5 x4 NVMe and twenty are hot-swap 2.5-inch SATA. They must be represented as separate storage pools.

Does it have M.2 storage?

Yes. Two internal M.2 sockets support Gen3 x4 NVMe or SATA 6Gb/s devices, depending on the configuration.

Does the base model support eight 600 W GPUs?

Do not assume that. ASRock Rack specifically describes related RF+ models as intended for 600 W accelerators. Use the current exact-model GPU QVL for the base 4U8G-TURIN2.

Is onboard 1GbE enough?

Usually not for production GPU data traffic. Use the OCP Gen5 x8 slot or qualified auxiliary adapter for the required storage, client or cluster network.

Is it an HGX server?

No. It is an eight-GPU PCIe server. It does not contain an HGX SXM baseboard or NVLink Switch fabric.

Verdict

ASRock Rack 4U8G-TURIN2 combines eight PCIe GPU positions with a well-provisioned dual-EPYC host and a useful 24-bay local storage layout. Four direct Gen5 NVMe bays, twenty SATA bays, 24 memory channels and an OCP network slot make it more balanced than accelerator-only chassis for data-rich workloads.

The buying discipline is exact variant control. The base model, RF+, RF and SUPER_4L versions are not interchangeable. GPU qualification, fan design and storage details must match the complete model string. The front drive pools also need separate four-NVMe and twenty-SATA limits in any configurator.

Configure the ASRock Rack 4U8G-TURIN2 around a currently qualified GPU, then validate the CPU, memory, storage, network, power and software plan as one system.

Technical sources

• ASRock Rack 4U8G-TURIN2 product page
• ASRock Rack 4U8G-TURIN2 and 4U10G-TURIN2 product brief
• ASRock Rack GPU support list
• ASRock Rack Q1 2026 GPU server guide
• AMD EPYC 9005 processor family

Specifications, support lists and firmware change. Confirm the full product suffix, GPUs, CPUs, DIMMs, storage backplane, OCP card, PCIe adapters, power feeds and software stack before purchase. This is a source-backed technical review, not a record of hands-on GPUMachines testing.