Blackwell · 8 GPUs per node · full FP64
8-GPU Blackwell servers that are equally at home training models and running double-precision science, built into clusters of one to 32 nodes.
Not every AI factory is only about AI. Universities, engineering firms and research institutes often need the same cluster to train a language model on Monday and run a fluid-dynamics or molecular simulation on Tuesday. HGX B200 is the Blackwell platform that does both well.
Each 8-GPU server carries 1.4 TB of HBM3e and keeps NVIDIA's full double-precision capability, rated at 296 TFLOPS of FP64 per board, while delivering the same 72 PFLOPS of FP8 AI performance as its Blackwell Ultra sibling. For mixed AI and simulation workloads, that balance is hard to beat.
We build B200 clusters from a single node up to a 32-node scalable unit, with the fabric, storage and scheduler (Slurm, Kubernetes or both) that research computing teams already know.
Blackwell Ultra cut FP64 to make room for AI. B200 kept it, so CFD, finite-element, climate and quantum-chemistry codes run at full speed on the same nodes as your AI training.
Fifth-generation NVLink at 1.8 TB/s per GPU and FP8 training with the second-generation Transformer Engine make multi-node training of large models efficient.
Slurm partitions, Kubernetes namespaces or Run:ai quotas let many research groups share one cluster with clear priorities and usage reporting for chargeback.
Around 53.6 kW for four nodes in a reference rack, a little below Blackwell Ultra, which can make the difference in an older machine room.
A typical node and cluster build, tailored to your codes and users.
| GPUs | 8 × NVIDIA Blackwell SXM per server, 1.4 TB HBM3e in total |
|---|---|
| FP4 Tensor Core | 144 | 72 PFLOPS per server |
| FP8 / FP16 Tensor Core | 72 / 36 PFLOPS per server (sparse) |
| FP64 | 296 TFLOPS per server |
| NVLink | Fifth generation, 1.8 TB/s per GPU, 14.4 TB/s per server |
| Network | Up to 400 Gb/s per GPU (0.8 TB/s per server) on InfiniBand or Spectrum-X |
| Cluster size | 1 to 32 nodes (up to 256 GPUs) per scalable unit, typically 4 nodes per rack |
| Software | Slurm and/or Kubernetes, Spack or EasyBuild software stacks, NVIDIA HPC SDK, NVIDIA AI Enterprise |
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.
What a B200 rack asks of your site.
| Rack power | Around 53.6 kW for four nodes (published reference figure) |
|---|---|
| Cooling | High-density air with containment, or direct liquid cooling |
| Storage | Parallel file system (for example Lustre, BeeGFS or a certified appliance) on its own network |
| Network | Leaf and spine fabric with room for the full scalable unit |
One accountable UK team from sizing to day-two operations
We look at your actual codes and models and agree the nodes, memory and fabric that suit them.
Power, cooling, floor and fibre checked, with a phased plan if the room needs upgrading.
Rack-level burn-in plus HPL, NCCL and storage benchmarks recorded for acceptance.
Slurm or Kubernetes configured with your identity system, partitions, quotas and user documentation.
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.
If double-precision simulation matters, B200. If the work is mainly LLM inference, reasoning and training, B300 offers more memory per GPU and twice the attention performance. Many sites mix both.
Usually, yes. We can integrate B200 nodes into an existing Slurm cluster and file system, with a fabric bridge if your current network is older.
Pricing and any programme support are confirmed on a written quote. Tell us about the project and funding route when you ask.
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.