Blackwell Ultra · 4 nodes · 32 GPUs
Four 8-GPU Blackwell Ultra servers, an 800 Gb/s fabric and fast storage in one tested rack: the most practical way to put serious reasoning and training capacity into a standard data hall.
Most organisations that want frontier-class AI capacity do not want to rebuild their data centre to get it. This rack is our answer: four 8-GPU NVIDIA HGX B300 servers, a GPU fabric running at 800 Gb/s per GPU, a storage tier and management network, delivered as one unit that has already been assembled, cabled and stress-tested before it reaches you.
Blackwell Ultra was designed for the way AI is used now. Reasoning models and agents think before they answer, which multiplies the number of tokens behind every reply. HGX B300 doubles attention performance over the first Blackwell generation and carries 288 GB of HBM3e per GPU, so long contexts and large models fit without being sliced across machines.
We treat the rack as the first step of a factory rather than the whole of it. It is built to a reference design, so a second and third rack join the same fabric without redesign when demand grows.
Reasoning models spend most of their time in attention. NVIDIA rates HGX B300 at twice the attention performance of HGX B200, which shows up as faster replies when many users are asking hard questions at once.
288 GB per GPU and 2.1 TB per server means a 400-billion-parameter model fits comfortably inside one node at FP8, with headroom for long conversations and large batches.
Inside each server, fifth-generation NVLink gives every GPU 1.8 TB/s to its neighbours. Between servers, one 800 Gb/s ConnectX-8 SuperNIC per GPU on a rail-optimised InfiniBand or Spectrum-X fabric keeps multi-node training efficient.
Air-cooled and liquid-cooled HGX B300 servers are both available, so the rack can go into a high-density air-cooled hall today or a direct-liquid-cooled suite where you have one.
The rack is one building block of a scalable unit of up to 32 nodes (256 GPUs). Additional racks plug into the same spine with the same cabling plan, so growth is a purchase order, not a project.
We burn in GPUs, memory, NVLink, the fabric and storage together and record the results, so you receive a rack with a known-good baseline rather than a pallet of boxes.
A typical build. We adjust CPUs, memory, storage, cooling and the fabric to your workloads and site.
| 算力 | 4 × HGX B300 8-GPU servers (our standard node is the air-cooled Compal SGX30-2, 10U; liquid-cooled nodes also available) |
|---|---|
| GPU | 32 × NVIDIA Blackwell Ultra SXM, 288 GB HBM3e each |
| GPU 显存 | 2.1 TB per server, 8.4 TB per rack |
| FP4 Tensor Core | 144 | 108 PFLOPS per server; 576 | 432 PFLOPS per rack |
| FP8 Tensor Core | 72 PFLOPS per server (sparse); 288 PFLOPS per rack |
| NVLink | Fifth generation, 1.8 TB/s per GPU, 14.4 TB/s per server |
| Host CPUs | Dual Intel Xeon 6 (or AMD EPYC, depending on the server maker) |
| GPU fabric | 8 × ConnectX-8 SuperNICs per server, up to 800 Gb/s each, on Quantum-X800 InfiniBand or Spectrum-X Ethernet |
| Storage and front-end | BlueField-3 DPUs or ConnectX NICs to a separate storage and service network; local PCIe Gen5 NVMe |
| 软件 | Ubuntu or RHEL, NVIDIA drivers and CUDA, NVIDIA AI Enterprise, Kubernetes or Slurm, optional Run:ai and Mission Control |
Figures are NVIDIA's published specifications for the platform. Where NVIDIA quotes two Tensor Core figures, the first assumes sparsity and the second is dense. The exact build (CPUs, memory, storage, cooling, networking) is agreed with you and confirmed on a written quote.
These are the numbers we check on site before quoting. If your hall cannot meet them, we will propose a colocation option or a different platform.
| Rack power | Around 60 kW for four HGX B300 nodes (published reference figure), plus network and storage |
|---|---|
| Per-server power | Around 14 kW for an 8-GPU Blackwell Ultra system (NVIDIA DGX B300 figure) |
| 冷却 | High-density air with hot-aisle containment, or direct liquid cooling with a coolant distribution unit |
| 机架 | Deep, high-load 48U to 52U rack; floor loading checked against the populated weight |
| Power feeds | Redundant three-phase supplies sized for the full load, with the PDU layout agreed in advance |
| 网络 | Space and power for leaf and spine switches if this is the first rack, and fibre routes to future racks |
One accountable UK team from sizing to day-two operations
We translate your users, models and training plans into tokens per second and GPU-hours, then show how many racks that needs now and in two years.
We survey power, cooling, floor and fibre, or shortlist UK colocation halls that can take 60 kW racks, before an order is placed.
Assembly, cabling, firmware alignment and multi-day stress tests at rack level, with a written acceptance report.
On-site installation, fabric bring-up, Kubernetes or Slurm, NVIDIA AI Enterprise and a first model served end to end with your team.
NVIDIA data-centre GPUs are allocated by NVIDIA and the server makers, and lead times are measured in months rather than weeks. Complete, tested systems usually get allocation sooner than loose parts, so we quote these platforms as built systems. Shipments outside the UK, and many advanced GPUs within it, need an end-user and end-use check under UK and US export rules before we can confirm supply.
Trade members get our engineers' time on sourcing and design questions, early stock and waitlist priority, and free technical support on every order.
Power and heat. Four HGX B300 nodes already draw around 60 kW, which is the practical ceiling for many air-cooled halls. With direct liquid cooling and enough power, denser layouts are possible and we will design for them.
Not quite. Each server is its own 8-GPU NVLink domain, and the four servers talk over the 800 Gb/s fabric. For training and most inference this scales very well. If a single model needs a larger NVLink domain, GB300 NVL72 joins 72 GPUs into one.
Yes. We can deliver a single HGX B300 node built to the same design, with the switch ports and cabling plan in place for the rest of the rack.
Pricing depends on the server maker, CPUs, memory, storage, fabric and cooling, and on allocation at the time. We quote these systems in writing after sizing and a site check.
Tell us what you want it to produce, how many people will use it and where it would live. An engineer will come back with a sized recommendation and an honest view on whether owning makes sense for you.