AI, rendering and computing
Professional graphics cards for AI
Professional graphics cards NVIDIA RTX PRO Blackwell with 16 to 48 GB of GDDR7 memory with ECC for AI, language models, rendering and scientific computing. Here you will find the RTX PRO 5000 (48 GB), RTX PRO 4500 (32 GB), RTX PRO 4000 (24 GB, also in a compact SFF version) and RTX PRO 2000 (16 GB) – energy-efficient low-profile versions with a power draw of 70 W fit even into small workstations and servers.
VRAM for LLMs: graphics memory determines how large a model will run on the card. 16 GB is enough for models of around 8 billion parameters, 24 GB for up to roughly 14 billion in 8-bit or around 30 billion in quantised form, 32 GB comfortably for models of around 30 billion parameters and 48 GB for quantised models of around 70 billion parameters with a shorter context. With several cards in one computer, their memory adds up.
Who it is for: companies and developers who want a GPU for AI in an existing workstation or server, as well as architects, designers and scientists. Ready-made builds can be found on the AI servers page, and we explain how to choose in the article AI server for business, work and home.
- 7 systemsfrom 40 290 Kč incl. VAT
- Pre-installedswitch it on and the server is ready to work
- Made to orderwe adjust the configuration to your needs
Server comparison
Prices excl. VAT, with VAT below. We will adjust the configuration to suit you.
| Server | GPU | VRAM | Processor | RAM | Power draw | Price excl. VAT |
|---|
VRAM = total graphics memory of all cards. Power draw at full load is approximate.
What the servers are for
- AI and LLMs
Language models in your company
An internal chatbot, assistant or working with company documents (RAG). Your data never leaves the company and you pay no monthly fees for cloud GPUs.
- Images, video, 3D
Generation and rendering
Image and video generation (FLUX.1, SDXL, ComfyUI), rendering in Blender and workloads that need lots of cores and memory.
- Bonus
Mining in idle time
When the server isn't running your workloads, it can mine cryptocurrencies on its graphics cards. See how much the cards earn on our graphics card profitability page.
AI server FAQ
Why a professional graphics card instead of a gaming one?
RTX PRO cards have ECC memory that corrects errors during long computations, certified drivers for professional applications and lower power consumption. They tend to be more compact and efficient, so more of them fit into one computer – and graphics memory is key for language models.
How much VRAM do I need for a language model?
As a rough guide: 16 GB for models of around 8 billion parameters, 24 GB for around 14 billion in 8-bit or 30 billion in quantised form, 32 GB comfortably for models of around 30 billion, and 48 GB for quantised models of around 70 billion parameters with a shorter context. The requirement grows with context length and the number of users.
Can I put several cards into one computer?
Yes, if the motherboard has enough slots and the computer has a sufficient power supply and cooling. For language models, the memory of the cards adds up. We will advise which combination suits your build, or prepare a complete AI server.
What are the SFF and low-profile versions for?
The RTX PRO 4000 SFF and RTX PRO 2000 are low-profile with a power draw of 70 W and need no auxiliary power. They fit into compact workstations and servers where a standard card would not fit or could not be powered.
Will you help with installing the card and software?
Yes. We can install the card in your computer and set up the drivers, CUDA and tools such as Ollama, vLLM or ComfyUI. The warranty is 24 months and you handle claims directly with us.
Discuss a configuration
Tell us which models or programs you want to run, how many users and what your budget is. We will propose a configuration and send you a quote.
- Phone+420 792 480 835
- Emailobchod@pcpraha.cz
- Shop and service centreRostovská 260/2b, Praha 10 · Mon–Fri 9–17