Most people arrive knowing the model, not the memory requirement. Start here and we will work backwards to the infrastructure.
~90 GB
Based on weights at the chosen precision plus working memory for context and concurrency. Real requirements vary with context length, batching and the serving framework. Treat this as a starting point for a conversation, not a specification.
A large-memory workstation card. It fits into a professional tower and lets you run or fine-tune substantial models on your own hardware.
Memory
96 GB
Deployment
Deskside workstation
£7,495
ex VAT
In stock
A very capable data centre GPU in card form, so it can be fitted into common rack servers rather than requiring a specialist chassis.
Memory
141 GB
Deployment
Rack server
£24,950
ex VAT
Low stock
The mainstream high-end training GPU. Sold as part of an eight-GPU server rather than as a card you fit yourself.
Memory
180 GB
Deployment
Rack server
Price on request
Available to order
A proven AMD accelerator with 192 GB of memory, frequently chosen for serving large language models economically.
Memory
192 GB
Deployment
Rack server
Price on request
Available to order
AMD's flagship accelerator. Very large memory per GPU means fewer GPUs are needed to hold a big model.
Memory
288 GB
Deployment
Datacentre
Price on request
Available to order
The GPU engine used inside eight-way AI servers. It is fitted into a server chassis by the manufacturer, not installed like a plug-in card.
Memory
2304 GB
Deployment
Datacentre
Price on request
Allocation only
Deployment classes shown on each card: Deskside workstation, Rack server, Datacentre.