GPUmachines

ASRock Rack 4U4G-TURIN/HPR Review: 4-GPU EPYC Server

ASRock Rack's 4U4G-TURIN/HPR combines one EPYC 9005/9004 CPU with up to four 600W PCIe GPUs, eight DDR5 slots and a 2+1 2700W PSU design.

ASRock Rack 4U4G-TURIN/HPR Review: 4-GPU EPYC Server

The ASRock Rack 4U4G-TURIN/HPR is a 4U single-socket AMD EPYC server built for as many as four 600 W double-width PCIe GPUs. It combines current EPYC 9005 and 9004 processor support with eight DDR5 DIMM slots, seven full-height expansion positions, two M.2 sockets and a 2+1 2,700 W power-supply arrangement.

That specification makes it a genuine GPU compute platform, not a general-purpose CPU server with a few spare PCIe slots. It suits organisations that need several independent accelerators in one node for AI inference, model fine-tuning, rendering, virtual workstations or scientific computing. It is not an HGX system: there is no fixed SXM baseboard or NVLink Switch fabric, and every GPU remains a qualified PCIe add-in card.

Configure the ASRock Rack 4U4G-TURIN/HPR on GPUMachines after checking the exact GPU, slot, power and operating-system combination. A four-GPU bill of materials needs more attention than the chassis headline suggests.

Technical review summary

ASRock Rack specifies support for one AMD EPYC 9005 or 9004 processor, including selected 3D V-Cache models and EPYC 97x4 parts, on socket SP5. The board accepts a CPU cTDP of up to 400 W. A single-socket design avoids cross-socket NUMA traffic and can simplify software licensing, but the chosen processor still has to feed four GPUs, networking and storage without becoming the host-side bottleneck.

The chassis can take four double-width, full-height, full-length GPUs rated up to 600 W each. An alternative layout supports as many as seven single-width cards. There is also one full-height Gen5 x8 position, although detailed lane sharing and the physical clearance around installed GPUs must be checked before treating every published slot as simultaneously usable.

System memory is the main compromise. Eight DDR5 DIMM sockets support RDIMM and RDIMM-3DS modules at up to 6,400 MT/s, subject to the CPU and qualified memory list. Current EPYC 9005 and 9004 processors expose twelve memory channels, so an eight-DIMM motherboard does not populate every available CPU channel. That can be acceptable for accelerator-led jobs, but buyers with heavy CPU preprocessing, large in-memory datasets or bandwidth-sensitive simulation codes should test the intended memory population.

Local storage is modest. The platform provides two M.2 Gen5 x4 sockets for 2280 or 22110 devices and a fixed cage for two 2.5-inch drives. It does not offer a bank of front hot-swap NVMe bays. Data-intensive deployments therefore need a deliberate external storage path, usually through a high-speed network adapter in one of the available expansion positions.

Dual 10GbE ports use a Broadcom BCM57416 controller, while an ASPEED AST2600 BMC supplies out-of-band management. The 2+1 power system uses 2,700 W 80 PLUS Titanium CRPS units. Four 600 W GPUs alone represent 2,400 W of device rating, before the CPU, memory, fans, NICs and drives are counted. The final combination must stay inside ASRock Rack's qualified power envelope rather than being approved from arithmetic alone.

Verified system specification

| Area | ASRock Rack 4U4G-TURIN/HPR specification | | --- | --- | | Form factor | 4U rackmount or tower-capable chassis | | Processor | 1 x AMD EPYC 9005 or 9004, socket SP5 | | CPU power support | Up to 400 W cTDP, subject to QVL | | Memory | 8 x DDR5 DIMM, 1DPC, RDIMM or RDIMM-3DS | | Published memory speed | Up to 6,400 MT/s, dependent on CPU and population | | Published module capacity | RDIMM up to 128 GB; RDIMM-3DS up to 256 GB | | Double-width GPU capacity | Up to 4 x full-height, full-length, double-width PCIe GPU | | GPU power class | Up to 600 W per qualified GPU | | Single-width capacity | Up to 7 x full-height, full-length cards | | Additional expansion | 1 x full-height, full-length PCIe Gen5 x8 position | | M.2 storage | 2 x Gen5 x4, supporting 2280 and 22110 devices | | Additional local storage | Fixed cage for 2 x 2.5-inch drives | | Onboard network | 2 x 10GbE RJ45 through Broadcom BCM57416 | | Management | Dedicated IPMI through ASPEED AST2600 BMC | | Cooling | 4 x middle 80 mm fans plus 2 x rear 80 mm GPU fans | | Power supplies | 2+1 x 2,700 W 80 PLUS Titanium CRPS | | Published dimensions | 710 x 430.4 x 174.8 mm with GPU fan section |

These are platform limits, not a promise that every CPU, DIMM, GPU, NIC and drive combination is qualified. The ordering revision, firmware, power cables, GPU brackets and operating-system support still need to be checked against the current vendor lists.

What the four-GPU layout is for

The 4U4G-TURIN/HPR is designed around standard PCIe accelerators. Each GPU is an independent device connected through the host PCIe topology, so the platform is well suited to workloads that can assign a model, render job, simulation or virtual desktop to a particular card.

Common uses include:

  • multi-model or multi-tenant AI inference, with services allocated to separate GPUs;
  • fine-tuning jobs that fit within the memory and communication limits of qualified PCIe cards;
  • retrieval, embedding, computer-vision and video-analysis pipelines;
  • GPU rendering and content-production workloads;
  • virtual workstation hosting where each user or group receives a physical or virtual GPU allocation;
  • CUDA or other accelerator-aware scientific applications that scale across PCIe devices; and
  • development environments that must mirror a larger production GPU estate without buying an eight-GPU node.

Four GPUs in 4U can also be operationally attractive. Compared with an eight-GPU chassis, the node has a smaller failure domain, lower acquisition cost and fewer accelerators taken offline during maintenance. Buyers can scale in four-GPU increments and allocate different nodes to different teams.

The trade-off is density. A rack filled with four-GPU 4U servers provides one GPU per rack unit before switches and storage are counted. Eight-GPU platforms may offer better rack density, while liquid-cooled or fixed-accelerator systems can go further. The best layout depends on site power, cooling, network topology and the size of each job, not only on GPU count.

PCIe GPUs are not HGX

The distinction matters when comparing this machine with the HGX server range. HGX systems use an NVIDIA baseboard with SXM GPUs, NVLink and NVLink Switch. The accelerators form a tightly connected scale-up domain intended for models and HPC jobs that communicate heavily between GPUs.

The 4U4G-TURIN/HPR uses ordinary PCIe add-in cards. Peer-to-peer transfers may be available for qualified combinations, but the topology and bandwidth are not equivalent to an HGX baseboard. Software should not assume that four PCIe cards behave as one large accelerator.

PCIe is often the better fit when:

  • workloads are independent or need limited inter-GPU communication;
  • professional graphics, virtual workstation or display-capable cards are required;
  • the buyer wants to mix accelerator classes supported by the vendor;
  • four GPUs are sufficient; or
  • acquisition cost and component replaceability matter more than maximum scale-up bandwidth.

Choose HGX when large distributed training or tightly coupled inference repeatedly moves data between all GPUs and can justify the higher platform, fabric and facility cost. For many production inference fleets, four independent PCIe GPUs are the more practical resource unit.

CPU choice and host-side balance

One SP5 processor has to support every accelerator, NIC, storage device and host task in this server. Core count is therefore only one part of CPU selection.

High-frequency EPYC models can suit GPU inference, rendering and engineering applications with serial or lightly threaded orchestration. Higher-core-count CPUs may be preferable when the node performs substantial data preparation, decompression, tokenisation, feature engineering, compilation, virtualisation or CPU-only services beside the accelerators. Large cache can matter for selected simulation and analytics workloads.

AMD's EPYC 9005 family extends the SP5 platform with Zen 5 and Zen 5c options. The wider family reaches much higher core counts than most four-GPU hosts need, so the largest processor is not automatically the best purchase. CPU licence cost, frequency, memory demand and the electrical allowance left for the GPUs should be considered together.

NUMA is simpler than on a dual-socket server because there is no second CPU memory domain. The PCIe tree can still affect device locality. Before production use, map every GPU, NIC and NVMe device with the operating-system topology tools and bind processes where the application benefits from it.

The current GPUMachines PCIe GPU server range includes single- and dual-socket alternatives. A dual-socket platform can provide more DIMM slots, aggregate CPU cores and I/O placement options, but it also introduces cross-socket traffic and a more involved memory policy.

Eight DIMM slots need a deliberate memory plan

Eight DDR5 slots are enough for a useful capacity, especially with high-density RDIMM or RDIMM-3DS modules, but they are fewer than the twelve memory channels available on one SP5 processor. This is not a minor line-item difference. Host memory bandwidth can affect how quickly CPUs prepare batches, feed accelerators, manage checkpoints and run services around the GPU workload.

Capacity should be based on the complete software path. GPU memory holds model weights and active data on the accelerator, while host RAM may hold datasets, caches, queues, preprocessing state, virtual machines and failover processes. A server with ample VRAM can still stall or swap if host memory is undersized.

Good practice is to:

1. start with the application's measured host-memory peak rather than a fixed ratio to GPU memory; 2. use a symmetric, vendor-qualified DIMM population; 3. leave capacity headroom for the operating system, drivers and monitoring; 4. verify the supported speed for the exact CPU and module rank; and 5. test CPU-to-GPU data delivery under the real batch size and data format.

A workload that depends on maximum SP5 memory-channel bandwidth may be better served by a board with twelve DIMM sockets per CPU. The 4U4G-TURIN/HPR favours GPU capacity and a simpler single-socket host over maximum host-memory expansion.

GPU selection and slot planning

ASRock Rack publishes support for four double-width cards up to 600 W each, but buyers should select from the current GPU support list for the exact chassis revision. Mechanical size, auxiliary power connector, firmware, fan policy and driver support all matter.

GPU decisions should include:

  • memory capacity per card for the target model, scene or simulation;
  • required numeric formats and software certification;
  • active or passive cooling orientation;
  • card width, length and power connector placement;
  • whether peer-to-peer communication is required;
  • partitioning or virtual-GPU licensing needs; and
  • the expected sustained rather than nameplate workload.

The seven single-width positions can be useful for lower-profile accelerators, video cards, network adapters or specialist I/O. They do not mean seven 600 W GPUs. Physical clearances and the system's approved power configuration remain the governing limits.

Treat the extra Gen5 x8 position as part of the full slot map. A high-speed NIC, DPU or storage adapter consumes space, lanes and power. The desired four-GPU layout should be drawn with every auxiliary card before purchase so that a network requirement does not displace a GPU or obstruct airflow.

Storage is deliberately limited

Two M.2 sockets and two fixed 2.5-inch positions are enough for mirrored boot devices, local software, logs and a modest scratch area. They are not a substitute for the front hot-swap NVMe tier found on data-rich GPU servers.

For inference services with models staged locally, the available drives may be adequate if model updates are controlled and the working set fits. Training, video analytics and simulation often need much more throughput or capacity. In those cases, plan a qualified high-speed NIC and a shared parallel, scale-out or NVMe storage service.

Check the intended M.2 devices carefully. Enterprise endurance, power-loss protection, thermal behaviour and operating-system support matter more than consumer peak-speed claims. ASRock Rack's detailed specification also identifies lane-sharing considerations around one M.2 path and an internal expansion position, so the final device map should be confirmed from the current manual.

The fixed 2.5-inch cage is useful for service drives, but fixed media is less convenient than front hot-swap storage. If local-drive serviceability is important, compare the platform with a GPU server that exposes front NVMe bays.

Networking for one node and for a cluster

The onboard dual 10GbE interface is suitable for management traffic, ordinary application access and smaller data transfers. It is rarely enough for a four-GPU node that reads large datasets, serves high-rate inference or joins a distributed training cluster.

A production design may add 100, 200, 400 or 800Gb/s networking, depending on the GPU generation, workload and storage system. Port speed alone does not determine results. The NIC's PCIe width and locality, switch capacity, optics, cables, routing, congestion control and storage endpoints all have to sustain the intended traffic.

Separate the traffic classes where scale justifies it:

  • out-of-band management through the BMC;
  • client or service traffic;
  • storage and checkpoint traffic; and
  • east-west GPU communication for multi-node jobs.

One adapter can carry several logical networks in a smaller deployment, but that choice should be explicit. For distributed AI, model how much communication crosses the network and whether the application can tolerate oversubscription. Four fast GPUs can expose an under-sized data path very quickly.

Power, cooling and rack deployment

The 2+1 2,700 W CRPS specification needs engineering review for a fully loaded configuration. Four GPUs rated at 600 W account for 2,400 W of device power. A supported EPYC CPU can add up to 400 W cTDP, and the motherboard, DIMMs, six high-pressure fans, NICs and drives add further load.

Those ratings must not simply be added and treated as measured wall power, but they show why the approved GPU list and vendor power rules matter. Confirm which PSU feeds, redundancy state and GPU limits apply to the selected build. A valid configuration may require a different GPU power class, CPU choice or operating limit.

The published chassis depth is 710 mm with the GPU fan section, and the server occupies 4U. Rack selection should allow room for rails, front and rear cables, power distribution and service access. Six 80 mm fans indicate that this is a data-centre platform; acoustic suitability for an office should not be assumed.

Facility planning should cover:

  • measured maximum and typical input power for the ordered build;
  • voltage, connector and PDU outlet requirements;
  • A/B feed design and the intended PSU redundancy mode;
  • rack-level power after switches and storage are included;
  • heat rejection at sustained GPU load;
  • cold-aisle temperature and airflow containment; and
  • the effect of a failed fan or PSU on allowed GPU operation.

Use the final tested server configuration in the rack planner. A generic barebone wattage or the sum of PSU nameplates is not an adequate facility figure.

Who should buy the 4U4G-TURIN/HPR?

This platform is a good candidate for:

  • AI teams running several independent inference services;
  • organisations fine-tuning models on up to four qualified PCIe GPUs;
  • render farms and engineering groups that need full-length professional cards;
  • service providers allocating GPUs to separate customers or virtual workstations;
  • laboratories that want a single-socket x86 host with modern Gen5 expansion; and
  • buyers scaling a fleet in manageable four-GPU nodes.

It is less suitable when the job needs an eight-GPU NVLink scale-up domain, many front hot-swap NVMe drives, all twelve SP5 memory channels, quiet office operation or very high GPU density per rack.

When not to buy it

Do not choose this server merely because four GPUs fit. Consider another platform if the application requires HGX-class GPU-to-GPU bandwidth, if every GPU must operate at 600 W without a vendor-confirmed power plan, or if the data pipeline depends on a large local NVMe bank.

A CPU-heavy application may also expose the eight-DIMM design. If preprocessing or simulation depends on maximum host-memory bandwidth, compare a twelve-DIMM single-socket server or a balanced dual-socket model. For one or two GPUs, a smaller chassis may use rack space and power more efficiently.

Configuration checklist

Before approving an order, record:

1. the exact CPU OPN, cTDP setting and firmware support; 2. the DIMM part number, capacity, rank, speed and eight-slot population; 3. the exact GPU model, quantity, power class, bracket and support-list status; 4. the complete PCIe slot map, including NICs, DPUs and storage adapters; 5. the M.2 and 2.5-inch drive plan, RAID or mirroring method and endurance requirement; 6. the network protocol, port speed, switch, optic or cable and traffic separation; 7. the PSU feed, redundancy mode, input voltage and rack PDU capacity; 8. the operating system, accelerator driver, runtime and application version; and 9. the acceptance tests for thermals, sustained power, GPU errors, storage and network throughput.

This checklist is more useful than selecting the largest component in every menu. A balanced four-GPU node keeps the accelerators supplied with data, remains serviceable and fits the actual rack envelope.

Frequently asked questions

How many GPUs fit in the ASRock Rack 4U4G-TURIN/HPR?

ASRock Rack specifies up to four full-height, full-length, double-width PCIe GPUs, with qualified models rated up to 600 W each. An alternative arrangement supports up to seven single-width cards.

Does it support NVIDIA HGX or SXM GPUs?

No. This is a PCIe GPU server. It does not contain an HGX baseboard, SXM modules or NVLink Switch fabric.

Which processors are supported?

One AMD EPYC 9005 or 9004 processor on socket SP5, including selected 3D V-Cache and EPYC 97x4 models. The published CPU cTDP limit is 400 W, subject to the current QVL.

How much system memory can be installed?

The motherboard has eight DDR5 DIMM sockets. ASRock Rack lists RDIMM modules up to 128 GB and RDIMM-3DS modules up to 256 GB, with speed up to 6,400 MT/s depending on CPU and population. Confirm current qualified modules before ordering.

Does the server use all twelve EPYC memory channels?

No. The board has eight DIMM sockets, while SP5 EPYC processors provide twelve memory channels. Memory-sensitive workloads should be tested with the intended eight-DIMM population.

How much local storage is available?

Two M.2 Gen5 x4 sockets and a fixed cage for two 2.5-inch drives. The platform does not provide a front bank of hot-swap NVMe bays.

Is onboard 10GbE enough for four GPUs?

It can be enough for management and lighter application traffic. Data-intensive inference, training or shared-storage use will usually need a faster qualified NIC and an appropriate switch and storage path.

Can four 600 W GPUs always run at full power?

Do not assume so from the slot count alone. Four such GPUs represent 2,400 W before host components are counted. The exact GPU, CPU, PSU redundancy and facility configuration must be approved against ASRock Rack's current support information.

Is it suitable for an office?

It is a 4U data-centre server with six 80 mm fans and high-power CRPS units. Plan for rack mounting, controlled airflow and substantial fan noise.

Verdict

The 4U4G-TURIN/HPR is a focused four-GPU PCIe platform. It offers current single-socket EPYC support, room for four large 600 W-class accelerators, modern Gen5 expansion and 2+1 Titanium power in a serviceable 4U format.

Its limits are equally clear: eight DIMM sockets, little local storage and no HGX interconnect. Those are reasonable trade-offs for independent or lightly coupled GPU jobs, provided the external storage, network and power plan are designed with the server rather than added afterwards.

Configure the ASRock Rack 4U4G-TURIN/HPR with the exact CPU, memory, GPU, storage and network requirements, then validate the completed bill of materials against the current vendor qualification lists.

Technical sources

Specifications, support lists and firmware change. Confirm the chassis revision, CPU, DIMMs, GPUs, slot map, drives, network adapters, power feeds and software stack before purchase. This is a source-backed technical review, not a record of hands-on GPUMachines testing.

← Back to blog