Graphics Card vs Data Centre GPU: What Is the Difference?
The differences between a graphics card for the PC and a data centre GPU for servers.
A graphics card is made for a PC or workstation, with its own fans and display outputs. A data centre GPU is made for servers: usually more and faster memory, a design for continuous operation, fast connections between GPUs and support for remote management. For gaming and local experimenting a graphics card is often enough; for AI at scale you look at data centre GPUs.
The key differences
| Aspect | Graphics card | Data centre GPU |
|---|---|---|
| Purpose | PC, workstation, gaming, creative work | Server, AI, compute in clouds |
| Memory | Usually less than data centre models | Often more and faster memory |
| Cooling | Own fans | Often passive, cooled by the server’s airflow |
| Connections | Mainly PCIe | Besides PCIe, often fast direct links between GPUs |
| Display outputs | Yes | Usually not |
| Management | Limited | Designed for management in servers and clusters |
| Availability | Shops and web shops | Through server suppliers and cloud services |
What does this mean for you?
If you want to try out a model or occasionally train something, a graphics card in your own machine can be a good start, provided the memory is sufficient. If you need a lot of capacity, want to link several GPUs or run continuously in production, renting data centre GPUs in the cloud is often more practical. Read renting or buying a GPU.
How do you compare them?
Our compare GPUs page lists graphics cards and data centre GPUs separately, with figures from the manufacturer. Compare within a segment and not across segments.
A practical example
A small studio wants to render video and occasionally try out a model. A graphics card in a workstation is then a logical choice: it suits the budget and the environment. A team that delivers a model to customers and runs it continuously is more likely to look at data centre GPUs in a cloud, because of memory, management and scalability.
A middle route is to test on a graphics card first and then scale up to rented data centre capacity once you know what you need.
What do ECC, MIG and NVLink mean for a data centre GPU?
The table above contains terms you will come across on product pages. What they mean:
- ECC memory: error correction that catches certain memory errors before they affect the result of a calculation. Data centre GPUs often have this; for graphics cards it differs per model.
- MIG (Multi-Instance GPU): an NVIDIA feature that splits one supported GPU into several smaller, separate units. Each unit has its own compute cores and memory, so that several users or tasks can share one card. Not every product supports this.
- NVLink: a fast connection between GPUs, intended for work where several cards together perform one task. Many models do not have this connection.
- Passive cooling: the card has no fans of its own and relies on the airflow of the server. Without that airflow, in an ordinary PC case, such a card can get too hot.
Whether you need these features depends on your work. If you only play games or try out a single model locally, you often have no need for them. For sharing between users or large models, they do play a role.
What should you look at before choosing or renting a GPU?
Use these questions as a checklist. A data centre GPU is not automatically the right choice, and neither is a graphics card.
- How much memory does your model or project need? Check your software’s documentation for how much video memory it asks for. See also what is HBM memory.
- Does the work run occasionally or continuously? With continuous use, cooling, remote management and support play a bigger role.
- Is there one user or are there several? With several users, a feature such as MIG, or simply multiple GPUs, can be useful.
- Do you need several GPUs for one task? Then check which connections the card supports and whether your server can handle them.
- Where is the hardware located? When renting, you also look at the location, especially if you work with personal data. See what is data residency.
- Which terms apply? Read the driver licence and the provider’s terms.
Can you use a graphics card in a server or data centre?
Technically, a graphics card can sometimes fit in a server, but that says nothing about what is permitted. NVIDIA has included a provision in the licence for the GeForce drivers: the software is not licensed for use in a data centre, with the exception of blockchain processing. This is the reason that companies computing on a large scale often choose products intended for that environment.
Other terms may apply for other manufacturers and for newer driver versions. So read the current licence and, if in doubt, ask the manufacturer or supplier. This is not legal advice; a lawyer can assess what applies to your situation. For an overview of specifications, see also GPU specifications explained.
Frequently asked questions
Can I put a graphics card in a server?
Sometimes, but manufacturers do not always design and support graphics cards for server use. For NVIDIA GeForce cards, the driver licence states that the software is not intended for data centre use (with the exception of blockchain processing). Check the terms and the documentation.
Is a data centre GPU automatically faster?
That depends on the product and the work. It is mainly designed for large, continuous workloads.
What is the difference between ECC and ordinary video memory?
ECC can detect and correct certain memory errors. Ordinary memory lacks that check. Whether a graphics card has ECC is stated in the specifications of the model.
Further reading
Last updated on
