NVIDIA 3XS HGX Custom AI Servers

AI Training & Inferencing Servers by 3XS Systems

3XS custom AI servers are based on NVIDIA-certified HGX designs. Powered by eight NVIDIA SXM GPUs, they provide data scientists and researchers with a powerful platform for scaling-out AI models. Designed for datacentre environments, 3XS AI servers are fine-tuned by our hardware engineers and data scientists for maximum performance and reliability.

We provide a custom software stack that means you spend more time on your model training and less time configuring drivers, libraries and APIs.

Buying HGX Servers Let's Chat

NVIDIA Grace Blackwell architecture showing CPU and GPU interconnect for AI workloads

NVIDIA HGX Platforms

There are multiple generations of HGX platform available, browse the different options in the tabs.

HGX Rubin Systems

The HGX Rubin NVL8 platform is powered by eight NVIDIA Rubin GPUs with 6th generation NVLink Switch, delivering 5.5x more NVFP4 FLOPS than HGX B200 for agentic and physical AI at scale. NVIDIA HGX Rubin NVL8 can be paired with either NVIDIA Vera CPU or x86-based CPU baseboards.

HGX Rubin NVL8 systems can be configured with ConnectX-9 NICs and BlueField-4 DPUs delivering Ethernet or InfiniBand connectivity at up to 800GB/s.

HGX Rubin NVL8 Specifications

HGX RUBIN NVL8
GPU	8x NVIDIA Rubin
Total GPU Memory \| Bandwidth	2.3TB HBM4 \| 160 TB/s
Performance	NVFP4 Inference: 400 PFLOPS NVFP4 Training: 280 PFLOPS* FP8/FP6 Training: 140 PFLOPS* INT8 Tensor Core: 2 PFLOPS* FP16 / BF16 Tensor Core: 32 PFLOPS* TF32 Tensor Core: 16 PFLOPS* FP64 Tensor Core: 264 TFLOPS
NVLink Switch System	NVLink Switch System
NVLink Bandwidth	28.8TB/s total bandwidth
Networking Bandwidth	1.6TB/s
Attention Performance	4x

*Dense specification.

HGX Rubin NVL8 Workloads

Agentic AI

The Rubin HGX architecture is specifically engineered to power AI agents that require multi-step reasoning, long-context understanding, and multi-modal, real-time decision-making.

Mixture-of-Experts Models

The Rubin HGX platform enables the training and inference of massive-scale MoE models. It is designed to train these models with 4x fewer GPUs compared to the Blackwell platform.

Massive-Context Inference

Rubin-based HGX systems can handle, process, and reason across million-token contexts, supporting complex software coding assistants and long-form video applications.

AI Factories

Rubin-based HGX systems serve as the compute engine for next-generation AI factories. They act as the building blocks for a unified, secure rack-scale architecture including networking and storage - functioning a single large GPU.

HGX Blackwell Systems

The HGX B300 and B200 platforms as powered by eight NVIDIA Blackwell GPUs with 5th generation NVLink Switch, delivering over 2x the performance of HGX H200 to power generative AI and AI factories. NVIDIA HGX B300 and B200 GPUs are paired with x86-based CPU baseboards.

HGX Blackwell-based systems can be configured with ConnectX-8 NICs and BlueField-3 DPUs delivering Ethernet or InfiniBand connectivity at up to 800GB/s.

HGX Blackwell Specifications

Key Specifications
GPU	HGX B300	HGX B200
Total GPU Memory \| Bandwidth	2.3TB HBM3e \| 160 TB/s	1.4TB HBM3e \| 160 TB/s
Performance	FP4 Tensor Core: 144 PFLOPS* FP8 Tensor Core: 72 PFLOPS* FP16 / BF16 Tensor Core: 36 PFLOPS* FP32 Tensor Core: 18 PFLOPS* FP64 Tensor core: 10 TFLOPS	FP4 Tensor Core: 144 PFLOPS* FP8 Tensor Core: 72 PFLOPS* FP16 / BF16 Tensor Core: 36 PFLOPS* FP32 Tensor Core: 18 PFLOPS* FP64 Tensor core: 296 TFLOPS
NVLink Switch System	5th generation	5th generation
NVLink Bandwidth	14.4TB/s total bandwidth	14.4TB/s total bandwidth
Networking Bandwidth	1.6TB/s	0.8TB/s
Attention Performance	2x	1x

*Dense specification.

HGX Blackwell Workloads

LLM Training and Fine-Tuning

The system provides faster, more efficient training of next-generation, parameter-dense models. It is particularly effective for training models that require deep attention chains and massive context windows, such as those over 1 million tokens.

Real-Time, Large-Scale Inference

Offering 144 PFLOPS of FP4 inference performance, these systems are specialised for high-traffic, real-time reasoning models and agentic AI. The ability to run 400B+ parameter models in GPU memory makes it highly efficient for complex AI services.

Industrial and Healthcare AI

Blackwell-based HGX systems support highly specialised workloads such as genomics, medical imaging and drug discovery for advanced healthcare applications; and complex digital twin environments for manufacturing and logistics.

Blackwell

Blackwell GPUs serve as the first compute engine designed specifically for AI factories. They act as the building blocks for a unified, secure rack-scale architecture including networking and storage - functioning as a single large GPU.

Agentic and Physical AI Made Simple

NVIDIA AI Enterprise, available at an additional cost with HGX systems, is a suite of frameworks and libraries, including NVIDIA NIM and NeMo microservices, that accelerate the development and deployment of AI projects. Leveraging pre-configured frameworks removes many of the manual tasks and complexity associated with software development, as each framework is tried, tested and optimised by NVIDIA. The less time spent developing, the greater the ROI on your AI hardware and data science investments.

Rather than trying to assemble thousands of co-dependent libraries and APIs from different authors when building your own AI applications, NVIDIA AI Enterprise removes this pain point by providing a full agentic AI software stack.

Workflow diagram of NVIDIA AI inference pipeline using models and microservices

Buying HGX Servers

HGX servers are available in different models from multiple brands, configured by our in-house expert engineers. Explore the key specs in the table below.


Platform	HGX Rublin NVL8	HGX 300	HGX B200
GPUs	8x NVIDIA Rubin	8x NVIDIA Blackwell Ultra	8x NVIDIA Blackwell
CPUs	2x Intel Xeon or AMD EPYC	2x Intel Xeon	2x Intel Xeon
Cooling	Liquid	Air	Air
RAM	TBC	Up to 32x DDR5	Up to 32x DDR5
Storage	TBC	12x NVMe SSDs	8x NVMe SSDs
Networking	NVIDIA ConnectX-9 NICs or Bluefield-4 DPUs up to 800Gb/s	NVIDIA ConnectX-8 NICs or Bluefield-3 DPUs up to 800Gb/s	NVIDIA ConnectX-7 NICs or Bluefield-3 DPUs up to 400Gb/s
Form Factor	2U	9U	10U
Get in touch	Register Interest	Configure	Configure

Frequently Asked Questions

What is NVIDIA HGX?

NVIDIA HGX is a family of GPU-accelerated servers that share the same underlying architecture as DGX appliances, but are configurable with different CPUs, memory, storage and software.

What is the difference between NVIDIA DGX and HGX?

The NVIDIA DGX and HGX are families of GPU-accelerated servers that share the same underlying architecture of multiple (typically eight) SXM GPUs communicating at high-speed with one another using the NVLink Switch bus. Whereas DGX appliances have a fixed hardware specification and software stack, some hardware elements, such as the CPUs, memory, storage of HGX servers can be configured to your requirements, and a custom software-stack installed.

What is the NVIDIA HGX Rubin NVL8?

The NVIDIA HGX Rubin NVL8 is a powerful AI supercomputer platform designed for agentic and physical AI at scale. It features eight NVIDIA Rubin GPUs and 800GB/s high-speed networking.

When is HGX Rubin NVL8 available to buy?

The NVIDIA HGX Rubin NVL8 is expected to be available to purchase in the second half of 2026. Register your interest here to be first in the queue.

How does the Rubin platform differ from Blackwell?

The next-generation Rubin platform includes the Rubin GPU (HBM4 memory), 6th gen NVLink Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet / Quantum X800 InfiniBand switch. In contrast the current generation Blackwell includes the Blackwell and Blackwell Ultra GPUs (HBM3e memory), 5th gen NVLink Switch, ConnectX-8 SuperNIC, BlueField-3 DPU, and Spectrum-4 Ethernet / Quantum-2 InfiniBand switch.

Almost every aspect of the latest Rubin platform features an uplift in performance over the previous generation, resulting in Rubin-based servers training and inferencing AI models significantly faster than Blackwell-based servers.

What workloads is the HGX Rubin NVL8 best suited for?

The DGX Rubin NVL8 is optimised for demanding AI workloads such as Agentic AI, Physical AI, large-scale Mixture-of-Experts (MoE) Models, Massive-Context Inference and next-generation AI factories.

What workloads are the HGX B300 and B200 best suited for?

The Blackwell-based HGX systems are optimised for demanding AI workloads such as large-scale deep learning training, generative AI (including LLMs and diffusion models), high-performance inference, data analytics, and scientific computing.

Can I buy the HGX systems in the UK?

Yes. Scan is an authorised NVIDIA Elite Partner and will be offering the HGX Rubin NVL8, alongside the HGX B300 and HGX B200 - all with full configuration, deployment, and support services. Scan is the only UK-based DGX Managed Services Provider (MSP) that enables businesses to access the power of DGX and HGX infrastructure without needing the skillset on-site to oversee the infrastructure. Instead, DGX MSPs provide full wraparound services to assist with anything from the initial deployment & configuration through to on-going services that encompass support, updates and overall platform management.

01204 474210

NVIDIA DGX Station GB300 – Desktop AI Supercomputer for Generative & Reasoning Models