ASRock Rack 2U4G-GENOA/M3 — 2U 4-GPU EPYC Barebone (PCIe 5.0, 1+1 2700 W)
Enterprise-grade platform, assembled and tested in the EU.
ASRock Rack 2U4G-GENOA/M3
Four full-size GPU in 2U, on one EPYC socket
The chassis we reach for when a four-card machine has to fit a rack that is already full.
A 2U rackmount built around a single SP5 socket that takes four full-length, triple-slot GPU on native PCIe 5.0 x16 — no switch, no lane sharing. Four hot-swap Gen5 NVMe bays, an OCP 3.0 slot for networking and a genuinely redundant 1+1 2700 W Platinum supply. It does in 2U what most of the market needs 4U to do.
Specification
| Component | Detail |
|---|---|
| Form factor | 2U rackmount, 850 × 438 × 87.5 mm |
| Processor | Single Socket SP5 (LGA 6096). AMD EPYC 9005 (BIOS update required), 9004 incl. 3D V-Cache, and 97x4. Up to 400 W TDP |
| Memory | 8 DIMM slots, 1DPC. DDR5 RDIMM / RDIMM-3DS to 4800 MHz. 128 GB per RDIMM, 256 GB per 3DS module |
| GPU slots | Rear: 1 × FHFL triple-slot PCIe 5.0 x16 + 1 × FHHL PCIe 5.0 x16. Internal: 3 × FHFL triple-slot PCIe 5.0 x16. Plus 1 × FHFL PCIe 5.0 x8 |
| GPU clearance | 312.4 × 137 × 62 mm rear, 312.4 × 130 × 62 mm internal |
| Storage | 4 × hot-swap 2.5" NVMe PCIe 5.0 x4 + 2 × hot-swap 2.5" SATA, 6-port backplane. 2 × M.2 PCIe 5.0 x4 internal |
| Networking | 1 × OCP NIC 3.0 (PCIe 5.0 x16), dedicated IPMI GLAN |
| Power | 1+1 CRPS, 2700 W at 200–240 V (1200 W at 110 V), 80 PLUS Platinum |
| Cooling | 3 × 80 × 38 mm + 1 × 80 × 80 mm PWM, front-to-back |
| Management | ASPEED AST2600 BMC, IPMI, mini DisplayPort, 2 × USB 3.2 Gen1 |
Read the power budget before you choose cards
The 1+1 supply means one module has to carry the whole machine, so the honest budget is 2700 W, not 5400 W. That single number decides which GPU this chassis is for.
| Component | Detail |
|---|---|
| 4 × 300 W class (e.g. RTX PRO 6000 Max-Q) | ~1200 W of GPU, ~1800 W total. Comfortable, and the configuration this chassis was designed around |
| 2 × 600 W class (Server / Workstation Edition) | ~1200 W of GPU, ~1800 W total. Also comfortable |
| 4 × 600 W class | ~2400 W of GPU, ~3000 W total. Over budget once a supply fails. Do not specify it |
Note also that the three internal x16 positions have no external cutout, so an add-in card fitted there has no reachable I/O ports. For compute cards that is irrelevant. For anything you need to plug into, use the rear slots.
Where the 8 DIMM slots bite
SP5 is a twelve-channel platform and this board brings out eight slots at one DIMM per channel. You therefore run eight of twelve channels, whatever you populate, and memory bandwidth lands around two thirds of what the CPU could do. For GPU work, where the cards hold the working set, that trade buys you the 2U height and is usually the right one. For a memory-bound CPU workload it is the wrong chassis and we will tell you so.
It also sets the ceiling: 1 TB with 128 GB RDIMMs, 2 TB with 256 GB 3DS. If you are sizing system memory at twice total VRAM — which is what we recommend — four 96 GB cards want 768 GB, and that needs 96 GB or larger modules.
Rack and cooling
Front-to-back airflow, rated for continuous 24/7 operation in a datacentre rack. Plan a dedicated feed sized for the real load: around 2 kW for a four-card Max-Q build. Cold aisle below 24 °C and real airflow through the rack, not just perforated doors — 2U at this power density is unforgiving of a cabinet that merely looks ventilated.
What is included
What is in the box: chassis, backplane, motherboard, the full fan wall and the power supplies listed above, pre-fitted and cabled. It is a complete, working platform less the parts you choose.
What is not: CPU, memory, drives, GPU and any add-in NIC. Those are specified with you and priced separately — that is the point of buying a barebone rather than a finished server.
What Kentino adds
Every barebone we ship is unpacked, assembled, powered up and tested before it leaves Prague: fan wall and PSU redundancy checked, BMC firmware brought up to the current release, IPMI configured and the network details handed over in writing. You get a machine that has already been switched on once by somebody who knows what it should sound like.
Tell us the workload instead of the part numbers and we will specify the CPU, memory and storage for you, build it, burn it in and ship it as a finished server. The same chassis, the same price for the chassis — we simply add the parts and the work.
Price shown is EUR 5,526 excluding VAT for the barebone as described, EXW Prague (Incoterms 2020). Freight, insurance and any destination duties are for the buyer’s account. Typical lead time 10–28 days. Image: ASRock Rack.
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.