GPUmachines

ASRock Rack 4U4G-GENOAX Review: 4-GPU EPYC Workstation

ASRock Rack's 4U4G-GENOAX is a single-socket AMD EPYC 4U workstation with four dual-slot PCIe Gen5 x16 positions, eight DIMMs and onboard 10 GbE.

ASRock Rack 4U4G-GENOAX Review: 4-GPU EPYC Workstation

The ASRock Rack 4U4G-GENOAX is a 4U rackmount workstation built around one AMD EPYC processor and direct PCIe Gen5 expansion. ASRock Rack lists four full-height, full-length dual-slot x16 positions or seven full-height, full-length single-slot x16 positions, making the platform relevant to GPU compute, rendering, visualisation and other accelerator-led workloads.

It is important to read those figures as alternative slot layouts, not as a promise that every card combination is qualified. GPU dimensions, cooling, auxiliary power and the 2100 W system power envelope still determine what can be installed safely. A production configuration should name the exact accelerator models and validate them against ASRock Rack's support information.

Single-socket AMD EPYC platform

The workstation uses one socket SP5 processor. ASRock Rack specifies support for AMD EPYC 9005 and 9004 families, including processors with AMD 3D V-Cache technology and selected 97x4 parts. A single-socket design can be attractive where the workload needs many cores and PCIe lanes but does not benefit from the cost, NUMA complexity or power draw of a second CPU.

Processor choice should match the application. Frequency-focused parts can suit interactive visualisation, simulation stages or commercial software licensed by core. Higher-core-count CPUs can improve rendering, compilation, preprocessing and heavily parallel CPU work. A GPU-led build still needs enough host performance to prepare data and keep its accelerators occupied.

Memory architecture

The 4U4G-GENOAX provides eight DDR5 DIMM slots at one DIMM per channel. It supports DDR5 RDIMM and RDIMM-3DS. This layout gives the single EPYC processor one populated slot for each memory channel when all eight positions are used.

The eight-slot limit makes memory planning important. Buyers should choose capacity with future expansion in mind because increasing memory later may require replacing rather than adding DIMMs. A balanced eight-DIMM population normally gives the processor its intended channel bandwidth. Final speed and capacity depend on the selected CPU, DIMMs and ASRock Rack's current QVL.

PCIe expansion and GPU suitability

The published expansion layout supports either four full-height, full-length dual-slot PCIe Gen5 x16 devices or seven full-height, full-length single-slot x16 devices. A further full-height, full-length Gen5 x8 position is listed. This gives the chassis flexibility for GPUs, network adapters, storage devices and specialised accelerators.

For a four-GPU configuration, check more than slot count. The cards must fit the physical clearances, receive the correct auxiliary power and operate within the chassis airflow and total electrical budget. High-power modern GPUs can exceed what a nominal four-slot layout can support when combined with a high-TDP EPYC processor. The final bill of materials should therefore be approved as a complete system.

The seven single-slot layout may suit lower-power accelerators, network interfaces, capture cards or mixed I/O builds. It should not be interpreted as seven double-width GPUs.

Storage

ASRock Rack lists two M.2 positions. One can operate as PCIe Gen5 x4 or provide connectivity for four SATA 6 Gb/s devices, while the other is PCIe Gen5 x4. These positions can support mirrored boot media or compact local storage, depending on the chosen arrangement.

This is not a front-bay storage server. GPU and workstation deployments that need large local datasets should define how additional storage will be attached, whether through supported internal devices, PCIe storage, external NVMe, NAS or a high-speed shared filesystem. Storage bandwidth should be sized from the workload rather than added after the accelerator choice.

Networking and management

Two 10 GbE RJ45 ports are provided through a Broadcom BCM57416 controller. A dedicated management interface and ASPEED AST2600 BMC provide remote-management capability. The onboard network is useful for general data and administration, but multi-GPU training or fast shared-storage access may require a higher-speed NVIDIA ConnectX adapter in an available PCIe slot.

Network planning should account for the slot consumed by the adapter, switch port, optic or cable, and the difference between management, storage and compute traffic. A four-GPU workstation can be held back by a slow data path even when local compute is ample.

Power and cooling

The chassis uses a redundant 1+1 pair of 2100 W 80 PLUS Platinum power supplies. It includes two fixed 80 x 56 mm PWM fans and two fixed 80 x 38 mm PWM fans. This is a rackmount cooling design, so noise levels and front-to-rear airflow will differ from a deskside workstation.

The PSU rating sets an upper platform boundary, not a target operating draw. CPU, GPU, memory, drives, network adapters, fans and conversion losses all consume part of the available budget. Four maximum-power GPUs may not be compatible with a given CPU and accessory combination. Facility planning and component qualification should use the completed build rather than the bare chassis description.

Suitable workloads

The 4U4G-GENOAX can fit:

  • GPU rendering and content creation;
  • scientific and engineering applications using up to four supported dual-slot accelerators;
  • model development, fine-tuning and inference;
  • VDI or visualisation workloads with qualified GPUs;
  • CPU-heavy preprocessing paired with GPU acceleration;
  • mixed PCIe workflows that need several x16 devices.

It is less appropriate when a workload requires NVLink-connected HGX GPUs, eight high-power accelerators, dense front hot-swap storage or dual host CPUs. Those requirements point to a different server class.

Configuration checklist

Start with the application and exact GPU model. Confirm card width, length, height, cooling direction, auxiliary power and total wattage. Then select an EPYC CPU, eight-DIMM memory population, boot design and any required storage or network adapter.

A complete quote should also define rack depth, rails, power feeds, expected draw, switch connectivity, operating system and accelerator software. For business-critical use, include firmware alignment, burn-in and workload acceptance testing.

Frequently asked questions

How many GPUs can the 4U4G-GENOAX hold?

The chassis provides four dual-slot PCIe Gen5 x16 positions or seven single-slot x16 positions. Actual GPU support depends on the exact cards, power, airflow and manufacturer qualification.

Is it a dual-socket server?

No. It uses one AMD EPYC 9005 or 9004 series processor on socket SP5.

How many DIMM slots are available?

There are eight DDR5 DIMM slots, supporting RDIMM and RDIMM-3DS.

Does it have front hot-swap drive bays?

The manufacturer's published feature list focuses on two internal M.2 positions and does not present this model as a front-bay storage server. Confirm any additional drive arrangement before ordering.

Is 10 GbE onboard?

Yes. Two 10 GbE RJ45 ports are provided through a Broadcom BCM57416 controller.

Source

Specifications were checked against the ASRock Rack 4U4G-GENOAX product page. Exact component support remains subject to the current CPU list, memory QVL, firmware and system qualification.

← Back to blog