If your data centre is being asked to support AI workloads, you've probably come across NVIDIA HGX. It's the eight-GPU platform behind most of the serious AI servers on the market, and it's a very different beast from the single-GPU servers most UK data centre teams are used to.
This guide explains what HGX actually is, how it compares to NVIDIA's DGX, and what to consider if you're buying one.
What Is NVIDIA HGX?
NVIDIA HGX is the reference platform that powers most of the AI servers sold today. It combines eight GPUs on a single baseboard, linked by NVLink and NVSwitch, with a software stack built for AI training and inference. Server makers including Lenovo, Dell, Supermicro and Gigabyte build HGX systems around this design, which is why HGX hardware runs in data centres worldwide.
NVIDIA defines the platform and certifies the baseboards, so buyers get a consistent, well-supported architecture whichever vendor they choose. The official NVIDIA HGX platform page sets out the current specifications and supported GPU generations. For UK data centre teams, HGX is usually the fastest route to large-scale accelerated compute without designing a bespoke GPU cluster from scratch.
How HGX Systems Are Built
An HGX system starts with a baseboard that carries eight GPUs, connected by NVLink. NVSwitch sits at the centre of that fabric, letting every GPU talk to every other GPU at full bandwidth. That interconnect is what separates HGX from simply slotting eight GPUs into a standard server: the communication path matters as much as the GPUs.
The baseboard then drops into a server chassis with CPUs, memory, storage and networking, and the finished system runs the NVIDIA software stack, including CUDA, the NGC catalogue and the NVIDIA AI Enterprise suite. HGX platforms are available in 8-GPU form and, for some generations, 4-GPU configurations.
HGX vs DGX: What's the Difference?
| Aspect | HGX | DGX |
|---|---|---|
| What it is | Reference platform for OEM-built servers | NVIDIA's own turnkey appliance |
| Who builds it | Lenovo, Dell, Supermicro, Gigabyte and others | NVIDIA |
| Customisation | Hardware choices made by the vendor | Fixed configuration |
| Software stack | NVIDIA stack, configured by the vendor | NVIDIA stack, fully integrated |
| Typical use | Custom AI servers at scale | Out-of-the-box AI appliance |
| Cost | Varies widely by vendor and specification | Premium, limited configuration |
The practical difference is flexibility. DGX is a fixed, fully integrated appliance, while HGX gives server makers the freedom to pair the eight-GPU baseboard with the CPUs, storage and chassis that fit a specific workload. For UK buyers, that usually means better price-performance and a wider choice of suppliers, at the cost of more configuration decisions.
NVIDIA HGX Generations: A100, H100 and H200
| Generation | GPU | Memory | Memory bandwidth | Architecture |
|---|---|---|---|---|
| A100 | A100 80GB SXM4 | 80GB HBM2e | 2.0TB/s | Ampere |
| H100 | H100 80GB SXM5 | 80GB HBM3 | 3.35TB/s | Hopper |
| H200 | H200 141GB SXM5 | 141GB HBM3e | 4.8TB/s | Hopper |
Each generation roughly doubles the memory capacity of the last. The H200 keeps the Hopper architecture of the H100 but swaps in 141GB of HBM3e per GPU, which matters for large language model training and inference, where weights and the KV cache have to fit in GPU memory. Eton sells both the H100 HGX SXM5 8-way board and the H200 141GB 8-way HGX baseboard, plus the HGX H100 and H200 HGX systems landing pages with the current range.
Why UK Data Centres Choose HGX
The appeal of HGX for UK operators comes down to four things. First, scale: an 8-GPU baseboard delivers the compute density needed for serious AI work without racking dozens of single-GPU servers. Second, consistency: because NVIDIA certifies the design, every HGX system behaves the same way, which simplifies driver, firmware and software management across a fleet.
Third, choice of supply. A multi-vendor market means UK buyers can shop for the best price and lead time rather than being tied to one supplier. Fourth, the software ecosystem: CUDA and the NGC catalogue are the default tools for AI development, and HGX hardware is where they run best. Our H200 NVL technical guide covers the PCIe alternative in more depth, and the H200 buyers guide walks through the GPU options for UK enterprises.
What Can HGX Power in a UK Data Centre?
HGX platforms earn their place in the data centre through workloads that need more parallelism than a handful of GPUs can offer. Large language model training is the headline use case: models with hundreds of billions of parameters are split across all eight GPUs, with NVLink carrying gradients between them at high speed. The H200's 141GB of HBM3e per GPU means larger models and longer context windows fit on a single node, which cuts training time and simplifies the software stack.
Inference is equally important. A single HGX node can serve multiple concurrent model instances, and the eight-GPU fabric handles the batch parallelism that keeps token generation latency low. Beyond generative AI, HGX systems accelerate scientific simulation, computational fluid dynamics, weather modelling and high-performance data analytics, which is why HPC teams in UK universities and research institutions deploy the same platforms.
There is also the A100 generation to consider. While H100 and H200 dominate new purchases, A100 8-way baseboards remain a cost-effective option for established workloads, and NVIDIA continues to support the architecture in its software stack. Eton's HGX range covers all three generations, so the right choice depends on your performance budget and the memory requirements of your models.
Buying HGX Systems in the UK: What to Consider
Power and cooling are the first questions for any HGX deployment. An eight-GPU system draws serious wattage, and the H200 baseboard is available in air-cooled and liquid-cooled variants, so the choice depends on your data centre's cooling infrastructure. Warranty and support come next: HGX kit is a significant investment, and a three-year warranty on new hardware gives you predictable protection.
Eton Technology supplies HGX hardware as new stock. The H200 141GB 8-way HGX baseboard is priced at GBP 209,675.00, and the H100 HGX SXM5 8-way board is available on request. Complete HGX systems, including Supermicro, Lenovo, Dell and Gigabyte servers, are listed in the NVIDIA HGX collection. Prices for full systems vary with CPU, memory and storage, so request a quote for your exact configuration.
Lead times matter too. Eight-GPU platforms are built to order, and availability fluctuates with NVIDIA's allocation of the latest GPUs, so it pays to talk to a supplier early in the planning cycle. A reputable UK reseller will confirm stock, warranty terms and delivery dates before you commit, and can advise on which generation gives you the best balance of price and performance for the models you plan to run.
Frequently Asked Questions
What does HGX stand for?
NVIDIA does not publish an official expansion. HGX is simply the name of its multi-GPU server platform, built around 8-GPU baseboards connected by NVLink and NVSwitch.
What is the difference between HGX and DGX?
DGX is NVIDIA's own fully integrated appliance. HGX is a reference design that server makers such as Lenovo, Dell and Supermicro build their own systems around.
How many GPUs does an HGX system have?
Standard HGX platforms carry eight GPUs on a single baseboard. NVIDIA also offers 4-GPU variants of some generations.
Is HGX the same as H100 or H200?
No. HGX is the platform, while H100 and H200 are the GPU generations that power it. An HGX system can be built around A100, H100 or H200 GPUs.
Can I buy an HGX baseboard on its own?
Yes. Eton Technology sells the NVIDIA H200 141GB 8-way HGX baseboard as new stock with a three-year warranty, ready to pair with a compatible server chassis.
How much does an NVIDIA HGX system cost in the UK?
The H200 141GB 8-way HGX baseboard is priced at GBP 209,675.00. Complete systems vary by vendor, CPU, memory and storage, so request pricing for your configuration.
Get in Touch
Not sure which HGX configuration fits your workload? Talk to the team at Eton Technology about baseboards, complete systems and lead times, and get a quote tailored to your data centre.
Call us on 0333 999 7768 or email info@etontechnology.com.
Alternatively, browse the NVIDIA HGX systems collection or use the contact form and we will come back to you within one working day.
