Supermicro SYS-822GS-NB3RT NVIDIA HGX B300 8-GPU AI Server – 8U, Dual Intel Xeon 6700P, Gold Series

The Supermicro SYS-822GS-NB3RT is an 8U air-cooled server with the NVIDIA HGX B300 8-GPU board – eight Blackwell Ultra B300 GPUs with 288 GB of HBM3e each (2.3 TB per system), fifth-generation NVLink and up to 108 PFLOPS of dense FP4. Eight NVIDIA ConnectX-8 SuperNICs are integrated on the board, so every GPU has its own 800 Gb/s port for scale-out InfiniBand (XDR) or Ethernet without filling PCIe slots.

It is the system Supermicro positions for large-scale AI inference serving: reasoning models, long context windows and mixture-of-experts LLMs that need the extra memory per GPU. It is also part of Supermicro's Gold Series: two pre-configured, pre-tested versions (-01-G2 and -02-G2) that currently reach European customers in about 6–12 weeks.

Explore AI & GPU Servers and compare SYS-A22GA-NBRT (HGX B200, Intel) with AS-8126GS-NB3RT (HGX B300, AMD).

Key features

• 8× NVIDIA B300 SXM (Blackwell Ultra) on HGX B300 – 2.3 TB HBM3e, NVLink 1.8 TB/s per GPU
• 8× NVIDIA ConnectX-8 SuperNICs on board – 8× 800G OSFP (1.6 TB/s AI fabric)
• Dual Intel Xeon 6700P – up to 86 cores per CPU, 350 W TDP
• 32 DIMM slots – up to 4 TB DDR5-6400 (1DPC) or 8 TB DDR5-6000 (2DPC)
• 8× hot-swap E1.S NVMe + 2× M.2 NVMe with RAID
• Gold Series presets ship faster: 6–12 weeks · 6× 6600 W Titanium PSUs (3+3)

Workloads and use cases

• Production LLM inference: 288 GB per GPU fits larger models and KV caches per GPU, and Blackwell Ultra doubles attention performance versus Blackwell.
• Reasoning and agentic AI: Dense FP4 (108 PFLOPS per system) and 2.3 TB of HBM3e for long-context, multi-step reasoning workloads.
• Fine-tuning and post-training: NVLink Switch connects all eight GPUs at 1.8 TB/s each for full-node fine-tuning, RLHF and distillation.
• Scale-out clusters: Built-in 8× 800G ConnectX-8 ports connect straight into Quantum-X800 InfiniBand or Spectrum-X Ethernet fabrics.

NVIDIA HGX B300 8-GPU performance

System-level figures published by NVIDIA for the eight-GPU board; real application performance depends on model, precision and software.

Gold Series presets – faster delivery

Supermicro Gold Series systems are pre-configured and pre-tested best-sellers with key components already installed, which shortens lead times. For HGX systems the GPU allocation still applies – current lead time to Europe is about 6–12 weeks.

SYS-822GS-NB3RT full specifications

Configuration options

Every HGX server is configured to order. Typical options are listed below; final compatibility and part numbers are confirmed in your quote.

Delivery and lead time

• Estimated lead time: 6–12 weeks for the Gold Series presets; custom configurations up to ~30 weeks depending on GPU allocation.
• Quote: written B2B quote excluding VAT within 1 business day, including the confirmed delivery window.
• Shipping: palletised freight to your data centre in all 27 EU countries; delivery to other countries on request and subject to export-control checks.
• Before shipping: optional burn-in test with report, firmware update and OS pre-installation.

Data-centre checklist

• Power: 6× 6600 W Titanium PSUs in 3+3 redundancy on 200–240 VAC – plan two independent feeds/PDUs. The exact power budget is part of the quote.
• Space: 950 mm deep 8U chassis – a rack at least 1200 mm deep.
• Weight: 119 kg net – use a lift for installation and check rack and floor load ratings.
• Cooling: front-to-back airflow, inlet air 10–35 °C; hot/cold-aisle containment recommended.
• Network: plan leaf switches, optics and cables for the GPU fabric (8× 800G per server) plus management and storage networks.
• Service: Supermicro lists this model as onsite-service-required – a manufacturer onsite service plan is included in every quote.

Buying from Mineshop.eu

• New systems sourced through European IT distribution, with manufacturer warranty and onsite service options.
• B2B invoicing excluding VAT, EU reverse charge for VAT-registered businesses, prices in EUR.
• Hardware specialists who help you size GPUs, memory, storage and networking for your models.
• One supplier for the complete stack: servers, GPUs, 400G/800G networking, server memory and NVMe storage.

Ordering process

1. Click Request a quote and tell us the configuration, quantity, delivery country and timeline.
2. A hardware specialist confirms the configuration with you – by email or a short call.
3. You receive a written quote within 1 business day, with lead time, service plan and payment terms.
4. You confirm with a purchase order; we handle export-control and end-user paperwork.
5. Your server is built, tested and delivered; optional rack integration and onsite installation.

Compare Supermicro HGX servers

Frequently asked questions

What is the price of the Supermicro SYS-822GS-NB3RT?

The SYS-822GS-NB3RT is quoted per configuration. HGX B300 pricing depends on GPU allocation, the Gold Series preset or custom build you choose, memory, storage, optics and service plan. You receive a written B2B quote, excluding VAT, within 1 business day.

What is the difference between SYS-822GS-NB3RT-01-G2 and -02-G2?

Both are pre-configured Gold Series versions with 2 TB DDR5-6400, a 960 GB M.2 boot drive and eight ConnectX-8 800G ports. The -01-G2 uses two Intel Xeon 6768P CPUs; the -02-G2 uses two Xeon 6776P CPUs and adds an NVIDIA BlueField-3 B3220 DPU for north-south networking and storage traffic.

How fast can I get an HGX B300 server?

The Gold Series presets currently have the shortest lead time – about 6–12 weeks to Europe. Custom configurations can take up to ~30 weeks depending on GPU allocation. Your quote states the confirmed lead time.

HGX B300 vs DGX B300 – what is the difference?

Both are built on NVIDIA's eight-GPU HGX B300 design. DGX B300 is NVIDIA's own branded system with a fixed bill of materials and NVIDIA's software and support bundle. The Supermicro SYS-822GS-NB3RT gives you the same GPU board with a choice of CPUs, memory, storage, DPUs and service level, which often means more flexibility on configuration, price and delivery.

What power and cooling does it need?

Six 6600 W Titanium PSUs in 3+3 redundancy on a 200–240 V supply (output is rated 5900–6600 W per PSU depending on input voltage), 8U in a rack at least 1200 mm deep (the chassis is 950 mm), a rack rated for 119 kg and 10–35 °C inlet air. We confirm the power budget and cords with your quote.

Should I wait for NVIDIA Rubin?

NVIDIA's HGX Rubin NVL8 platform is ramping in 2026 and will offer more performance. If you need production capacity in the coming months, a Gold Series HGX B300 is usually the faster path; we can quote both so you can compare price and delivery.

NVIDIA, HGX, NVLink, ConnectX and BlueField are trademarks of NVIDIA Corporation. Supermicro is a trademark of Super Micro Computer, Inc. Specifications from the manufacturers; subject to change without notice.