ASRock Rack 4U10G-TURIN2/RF 4U Dual Socket SP5 AI GPU Server
ASRock Rack 4U10G-TURIN2/RF 4U Dual Socket SP5 AI GPU Server
The ASRock Rack 4U10G-TURIN2/RF is a 4U rackmount barebone server built for dense, air-cooled accelerator deployments. It pairs dual AMD EPYC 9005 (Turin) or 9004 SP5 processors with 24 DDR5 DIMM slots, ten full-height double-width PCIe 5.0 x16 slots behind a PEX89104 PCIe switch, and sixteen hot-swap PCIe 5.0 NVMe bays, all powered by 3+1 redundant 3200 W 80 PLUS Titanium supplies. The extra rack units over a 2U chassis provide airflow headroom rated for 600 W accelerators and room for ten double-width cards, with front-to-back cooling from five mid-chassis and five rear hot-swap 80 mm fans.
As a barebone platform, the chassis, motherboard, backplanes, cooling and power are supplied ready for configuration, while CPUs, memory, GPUs, storage and NICs are configured to order and burn-in tested before shipping. The 867 mm (34.1 in) depth requires a rack that can accommodate it, and the 3200 W output requires 200–240 V input; on 100–127 V each supply is limited to 1000 W.
Key Features
| Brand | ASRock Rack |
| Model | 4U10G-TURIN2/RF |
| Form Factor | 4U rackmount |
| Depth | 34.1 in (867 mm) |
| Processor Support | Dual AMD EPYC 9005 (Turin) / 9004, SP5 (LGA 6096) |
| Max TDP per Socket | Up to 500 W |
| DIMM Slots | 24 (12+12, one DIMM per channel) |
| Max Memory | 6 TB DDR5 |
| Supported Modules | 128 GB RDIMMs or 256 GB RDIMM-3DS |
| Max Speed | 6400 MT/s |
| PCIe Switch | PEX89104 |
| PCIe Slots | 10 FHFL dual-slot x16 (PCIe 5.0), 1 CPU-direct x16, 2 front-accessible FHHL x16 |
| Drive Bays | 16 × 2.5" U.2 hot-swap NVMe (PCIe 5.0 x4); optional 24 × 2.5" SATA/SAS with RAID/HBA card |
| Fans | 5 mid-chassis + 5 rear hot-swap 80 mm (push-pull) |
| GPU Support | Up to 600 W accelerators |
| PSU | 3+1 redundant 3200 W 80 PLUS Titanium CRPS |
| Input Voltage | 200–240 V for 3200 W; 100–127 V limited to 1000 W per supply |
| BMC | ASPEED AST2600 with dedicated IPMI LAN port |
| OCP | OCP NIC 3.0 (PCIe 5.0 x16) |
Highlights
- AI infrastructure teams standing up on-prem training or fine-tuning clusters where eight-GPU nodes are the unit of scale and cloud GPU spend has become the bottleneck.
- Enterprise and regulated-industry IT (finance, healthcare, legal, defense-adjacent) that need model weights and datasets to stay inside the building.
- Research universities and national labs that buy in node-sized increments and need a platform that accepts the widest range of accelerator, NIC and storage cards without a fork-lift chassis change.
- Managed service providers and inference hosts that would rather pay once for capacity than per-token, and need redundant power and hot-swap fans for 24/7 operation.
- HPC and rendering groups that want a general-purpose EPYC node with enormous PCIe and NVMe fan-out — GPU-optional, since the ten double-width slots also take high-end compute, FPGA or capture cards.
Compatibility
Supports dual AMD EPYC 9005 (Turin) or 9004 processors in the SP5 (LGA 6096) socket, including 3D V-Cache parts and the 97x4 series. Memory compatibility covers 128 GB RDIMMs and 256 GB RDIMM-3DS modules at up to 6400 MT/s across 24 DIMM slots. The ten FHFL dual-slot PCIe 5.0 x16 slots accept up to 600 W accelerators, and the OCP NIC 3.0 slot supports PCIe 5.0 x16 mezzanine cards.
Warranty
Extended 3-Year Warranty
What's included
Not included unless configured: CPUs, CPU heatsinks, DDR5 RDIMMs, GPUs and GPU power cables, NVMe/SATA/SAS drives, OCP or PCIe network cards, RAID/HBA cards, operating system and software stack.
Frequently asked questions
|