Sizing, hardware and networking for training, fine-tuning and serving large language models — specified around your model, context length and user count.
The model you intend to run decides memory, and memory decides the GPU. As a guide, allow roughly 2 GB of GPU memory per billion parameters at 16-bit precision, plus space for context.
Training needs interconnect bandwidth between GPUs. Inference needs memory capacity and memory bandwidth. Buying the wrong one is the most common and most expensive mistake.
Multi-node training is limited by the fabric. We specify InfiniBand or Ethernet fabrics appropriate to the cluster size.
Rent while you experiment, buy once utilisation is steady. We can model the crossover point for your workload.
Departmental inference, team fine-tuning, and multi-node training reference builds, all quotable as complete projects.
Container runtime, schedulers, inference servers and monitoring, configured before delivery.
A rack server holding four GPUs. It suits a comms room or colocation rack and serves models for a whole team or application.
Memory
Configurable
Deployment
Tower server
From £42,500
ex VAT
Available to order
A large training server. It needs three-phase power and proper cooling, so it belongs in a data centre or a purpose-built room.
Memory
Configurable
Deployment
Rack server
Price on request
Available to order
A previously deployed training server, fully tested and warranted. A cost-effective way to obtain serious training capacity.
Memory
Configurable
Deployment
Rack server
£58,500
ex VAT
Low stock
An entire rack delivered as one computer. It is bought for frontier model training and very large inference platforms, and requires liquid cooling and high-capacity power.
Memory
20736 GB
Deployment
Datacentre
Price on request
Allocation only