Supermicro AS -8125GS-TNHR 8U barebone server, dual SP5, HGX H100 8-GPU
An 8U barebone server supporting dual AMD EPYC 9004 CPUs, eight NVIDIA HGX H100 GPUs with NVLink and NVSwitch, 24 DDR5 slots for up to 6 TB memory, and 18 hot-swap 2.5" NVMe/SATA bays
- ProcessorAMD EPYC 9004 Series
- Graphics processorSupported GPU NVIDIA SXM: HGX H100 8-GPU (80 GB), Up to 8 do
- Installed RAMNo Memory Installed
- Form factor8U Rackmount
- Processor socketDual Socket SP5
- Target workloadAI/ML/HPC
- Designed for AI, ML and HPC workloads
- Dual SP5 sockets support AMD EPYC 9004 series CPUs up to 400W TDP
- 24 DDR5 slots accept 4800MHz ECC RDIMM or LRDIMM up to 6TB total
- HGX H100 8-GPU with NVLink and NVSwitch for high-speed GPU communication
- 18 hot-swap 2.5-inch NVMe and SATA drive bays for dense storage
HGX H100 8-GPU AI server
The Supermicro AS -8125GS-TNHR is an 8U barebone server built for AI, ML and HPC workloads. It supports dual AMD EPYC 9004 Series processors on SP5 sockets and integrates the NVIDIA HGX H100 8-GPU platform with 80GB per GPU. Up to eight double-width GPUs connect through PCIe 5.0 x16 CPU-to-GPU links and NVIDIA NVLink with NVSwitch for high-speed GPU communication.
Scale-out clusters, not small labs
This system suits organisations deploying large-scale training or inference clusters that need dense GPU compute and direct RDMA networking at a 1:1 GPU ratio. It is not for buyers who need a compact workstation, a single-GPU development box or a platform that fits in a standard tower chassis. Teams without rack power and cooling infrastructure should evaluate smaller form factors.
Eight SXM GPUs on HGX H100 baseboard
The defining specification is the integrated NVIDIA HGX H100 8-GPU baseboard that locks the system to eight SXM5 GPUs with NVLink and NVSwitch fabric. Memory capacity reaches 6TB across 24 DDR5 slots at 4800MHz using ECC RDIMM or LRDIMM modules. Storage density provides 18 hot-swap 2.5" NVMe or SATA bays for local datasets.
How it compares
Measured against everything we stock in Barebone Servers & DIY Kits, at today’s prices. Recalculated every night.
- Price $17,666.99
- Pricier than about 90% of the 532 comparable items (median $2,503.49).
New to this? Buying server hardware with crypto — what to check, what it costs to pay, and what is never asked for.
Specifications
| Product | |
|---|---|
| Brand name | SUPERMICRO |
| Model name | AS -8125GS-TNHR |
| Part number | AS -8125GS-TNHR |
| Target workload | AI/ML/HPC |
| Processor | |
| Processor | AMD EPYC 9004 Series |
| Processor socket | Dual Socket SP5 |
| Processor details | Supports up to 400 W TDP CPUs (Air Cooled) |
| Memory | |
| Installed RAM | No Memory Installed |
| RAM slots | 24 × 288Pin |
| Maximum supported memory | Up to 6 TB: 24 × 256 GB DRAM |
| Supported memory | 4800 MHz ECC DDR5 RDIMM/LRDIMM |
| Storage | |
| 2.5" bays | 18 × 2.5 in hot-swap NVMe/SATA drive bays |
| Graphics | |
| Graphics processor | Supported GPU NVIDIA SXM: HGX H100 8-GPU (80 GB), Up to 8 double-width GPU(s) CPU-GPU Interconnect: PCIe 5.0 x16 CPU-to-GPU Interconnect GPU-GPU Interconnect: NVIDIA NVLink with NVSwitch |
| Size & build | |
| Form factor | 8U Rackmount |
| Product dimensions | 14 in × 17.2 in × 33.2 in |
| Shipping weight | 102.06 kg |
| Package size | 1300 × 749 × 699 mm |
| Environment & compliance | |
| Temperature range | Operating Temperature: 10 °C to 35 °C (50 °F to 95 °F) Non-operating Temperature: -40 °C to 60 °C (-40 °F to 140 °F) |
| Humidity range | Operating Relative Humidity: 8% to 90% (non-condensing) Non-operating Relative Humidity: 5% to 95% (non-condensing) |
Questions about this item
Which AMD EPYC 9004 Series processors fit the dual SP5 sockets and 400W TDP limit?
The dual SP5 sockets accept AMD EPYC 9004 Series CPUs rated up to 400W TDP with air cooling. You must supply two matching processors before the system will POST.
Will my existing 4800MHz ECC DDR5 RDIMM or LRDIMM modules populate the 24 slots and reach the 6TB maximum?
Yes. The 24 x 288-pin slots support 4800MHz ECC DDR5 RDIMM or LRDIMM modules; filling all slots with 256GB DIMMs reaches the 6TB limit.
How do I confirm the HGX H100 8-GPU complex is linked at full speed after installation?
Verify each GPU shows a PCIe 5.0 x16 link to the CPUs and that NVLink with NVSwitch reports full bandwidth between all eight GPUs in the system firmware or NVIDIA tools.
More in Barebone Servers & DIY Kits
Similar items
ASUS RS720A-E13-RS8G-32KW10G-EPYC-9554 barebone server EPYC 9554 64-core 3.10 GHz 256GB DDR5 960GB SSD
- EPYC 9005 Series
- 2U Rackmount
GIGABYTE R184-S90-AAV1 barebone server 1U dual Xeon 6 4-bay Gen5 NVMe/SATA/SAS-4 32 DIMM
- Xeon 6 Processors,…
- 1U Rackmount
DELL PowerEdge R960 barebone server 16 x 2.5-inch backplanes LGA 3647
- Rack
- LGA 3647





