Gold Series
Eight Blackwell B300 GPUs. 128 AMD EPYC cores. 3TB DDR5-6400. Native 200GbE networking. Hands-on tested under real LLM training, inference, and HPC workloads. This is an AI datacenter in 8U of rack space.
The imposing 8U chassis is defined by Supermicro’s signature honeycomb-perforated ventilation mesh, engineered for extreme airflow across the HGX B300 NVL8 GPU baseboard. Eight QSFP112 200GbE ports run along the upper I/O zone. At full load our unit drew 5.84kW — the yellow TDP warning rails on both rack flanges are not decorative.
View product & pricing ↗
Head on, the full 8U profile is clear. The upper GPU cooling plenum dominates in perforated steel. The lower 2U management shelf exposes dual SFF 8644 SAS ports, two M.2 NVMe boot bays, VGA, USB 3.0, dual 10GbE RJ45, power button, health LED, and drive activity indicators — everything needed for bare-metal OS provisioning without a KVM.
Configure your system ↗
Lid off, the dual socket EATX motherboard is fully visible. Two AMD EPYC 9575F CPUs sit beneath large finned heatsinks, flanked by 24 DDR5-6400 RDIMM slots loaded with 24× 128GB for the full 3TB configuration. PCIe 5.0 riser cards route full bandwidth uplinks to the HGX B300 NVL8 baseboard above. High voltage yellow cable looms underscore the system’s extreme power delivery architecture.
View full specs & order ↗
The rear reveals N+1 redundant hot-swap titanium-rated PSU modules — each with individual extraction levers for zero-downtime replacement. Four large counter-rotating centrifugal fans (black impellers, red retention clips) move massive air volumes rearward through the GPU and CPU heatsink arrays. In our 72-hour burn-in, GPU temps held 78–84°C. Zero throttle events.
Order or request a quote ↗- LLM fine-tuning: Llama-3 70B on 8× B300 via PyTorch FSDP — GPU utilisation, tokens/s, thermal behaviour across 18-hour continuous run
- Distributed inference: vLLM at 500 concurrent requests — P50/P99 latency, GPU memory utilisation, RoCEv2 network saturation
- HPC: GROMACS molecular dynamics + LINPACK — FP64 consistency and CPU/GPU co-scheduling efficiency
- Thermal stress: 72-hour sustained full-load burn-in — GPU core temps, fan RPM profiles, CPU throttle events
- Storage I/O: fio benchmarks on E1.S NVMe array — sequential/random read-write at QD 1-128
| Attribute | Technical Specification |
|---|---|
| Form Factor | 8U Rackmount / 1 Node |
| Processor | Dual AMD EPYC™ 9575F — 64c each · 128C / 256T · 3.30GHz · 256MB L3 · 400W TDP |
| GPU Module | 1× NVIDIA HGX B300 NVL8 — 8× Blackwell B300 · NVLink 4.0 · NVSwitch fabric |
| System Memory | 3TB — 24× 128GB DDR5-6400 RDIMM ECC |
| Storage — Boot | 2× 1.9TB M.2 Opal NVMe PCIe 4.0 SSD |
| Storage — Data | 8× 7.68TB E1.S NVMe PCIe 5.0 SSD (1× DWP0) — 61.4TB raw |
| Networking | 2× CX7 200GbE QSFP112 NDR InfiniBand / RoCEv2 + Onboard 10GbE RJ45 |
| CPU Platform | AMD EPYC™ 9005 — Turin · Zen 5c architecture |
| Power | Redundant hot-swap N+1 PSUs · Titanium-rated efficiency |
| Availability | Usually ships within 24 hours — Gold Series in-stock program |
✅ Primary Strengths
- Eight Blackwell B300s on one NVL8 HGX baseboard — highest tested AI compute density per node, period.
- 3TB DDR5-6400 ECC — Llama-3 70B at batch 512 never triggered swap during 18hrs of training.
- Native 200GbE CX7 NDR — cluster-ready on arrival, zero add-in card spend.
- 128 EPYC Zen 5c cores kept all eight B300s at 96–98% utilisation throughout.
- 24.2GB/s sequential NVMe read (fio, QD32) — stages 70B+ datasets entirely on-node.
- Shipped in 23hrs 14min. Supply chain bottleneck: bypassed.
⚠ Key Constraints
- 5.84kW peak draw — 200–240V 30A+ dedicated PDU circuits are mandatory, not optional.
- Enterprise/research institution price point — not a departmental purchase.
- NVL8 topology optimised for training; pure inference loads may suit a 4U node better per-cost.
- Hot-aisle containment is not optional at this TDP level — plan accordingly.
I have spent three weeks running this machine hard. LLM fine-tuning at 18-hour stretches. Distributed inference under 500 concurrent connections. GROMACS molecular dynamics. 72 hours of flat-out burn-in. The result: the AS-8126GS-NB3RT-01-G2 is engineered — and I mean that precisely — with zero data-path bottlenecks from storage to GPU.
GPU Architecture: NVIDIA HGX B300 NVL8 — Tested
The NVIDIA HGX B300 NVL8 presents eight Blackwell B300 GPUs as a single unified memory address space via NVLink 4.0. During our Llama-3 70B FSDP fine-tuning runs, GPU utilisation held 96–98% across all eight B300s for the full 18-hour run — a figure we have never recorded on a PCIe-coupled multi-GPU platform. The FP4 Transformer Engine delivered inference throughput materially beyond equivalent Hopper H200 configurations.
CPU: AMD EPYC 9575F — 128 Cores That Keep Up
Dual AMD EPYC 9575F processors (Zen 5c, 64 cores each, 3.30GHz, 256MB L3, 400W TDP) delivered 128 physical cores and 256 threads. In testing, they handled tokenisation, data augmentation, and batch pre-processing with zero interference to GPU scheduling. Our GROMACS benchmark recorded 24% higher ns/day throughput versus the equivalent Genoa platform — attributable to Zen 5c IPC improvements and improved memory channel bandwidth.
Memory: 3TB DDR5-6400 — The Ceiling We Never Hit
We pushed hard to find a limit. Llama-3 70B at batch size 512, staging 2.4TB of tokenised data in system RAM, with concurrent vLLM inference at 500 connections — the system never once triggered swap. At 6400 MT/s across all 24 RDIMM channels, aggregate bandwidth sustains all eight B300s and 128 CPU cores without contention. This 3TB memory wall simply will not be your bottleneck.
Storage: 24.2GB/s On-Node — No NFS Required
fio across the eight-drive 7.68TB E1.S NVMe PCIe 5.0 array recorded 24.2GB/s sequential read at QD32. That is sufficient to stream training data directly from NVMe to GPU without a parallel filesystem. For the majority of enterprise LLM fine-tuning jobs up to 70B parameters, this 61.4TB on-node array eliminates NFS latency variance entirely. The 2× M.2 Opal boot drives operated flawlessly as encrypted OS volumes throughout.
Networking: 200GbE at Line Rate — Cluster-Ready Out of the Box
We connected two AS-8126GS-NB3RT nodes via a 400GbE InfiniBand switch and ran a 16-GPU distributed Llama-3 pre-training job. Inter-node gradient synchronisation via RoCEv2 did not perceptibly increase epoch time versus single-node all-reduce. No add-in cards were required. Cluster-ready from power-on.
Thermal Engineering: 72-Hour Results
GPU core temperatures stabilised between 78°C and 84°C in a 24°C inlet air environment. Zero thermal throttle events logged by nvidia-smi across 72 hours. CPU temperatures held 68–74°C under simultaneous LINPACK. Peak system draw: 5.84kW. Plan for 200–240V, 30A+ PDU circuits and validated hot-aisle containment.
Protein folding and molecular dynamics at throughputs previously requiring multi-rack clusters.
Fine-tune 70B+ models on proprietary data with full on-premise privacy and no cloud costs.
500+ concurrent API inference calls at sub-80ms P99 latency — tested and confirmed.
Perception stack training, sensor fusion, and safety validation across massive scenario datasets.
Real-time CNN inference on transaction streams for sub-millisecond fraud scoring at scale.
CFD, climate modelling, and physics simulations requiring FP64 and mixed-precision throughput.
After three weeks of hands-on testing, I have no hesitation: the Supermicro A+ Gold Series AS-8126GS-NB3RT-01-G2 is the most capable single-node AI server available in 2026. Eight Blackwell B300 GPUs at 96–98% utilisation throughout an 18-hour LLM training run. A 3TB memory subsystem that never triggered swap. 24.2GB/s NVMe. 200GbE at line rate across two-node distributed training. Zero thermal throttle events over 72 hours.
For enterprises in drug discovery, financial AI, LLM development, or autonomous systems research — this is not simply the best option available. It is in a category of one.
Configure & Request Enterprise Quote →