AI workstation for local language models
For running language models, generating images and training your own models, graphics card memory matters most. It decides how large a model fits entirely on the GPU, and that determines speed.
What drives performance
VRAM decides model size
A language model runs fastest when it fits entirely in the card's memory. An RTX 5090 with 32 GB of VRAM can run models of around 30 billion parameters in popular quantised versions. Larger models require several cards or a professional RTX PRO card with more memory.
Image and video generation
16 GB of VRAM is enough for Stable Diffusion and Flux in ComfyUI, but newer image and video models run clearly better on cards with 24-32 GB. More memory also allows higher resolutions and several models at once.
Training and fine-tuning
Fine-tuning, for example with LoRA, and training need more memory than inference. For such workloads we recommend the Threadripper platform, which holds several graphics cards, lots of RAM and fast NVMe drives for datasets, together with a power supply with plenty of headroom.
System and software
NVIDIA cards are the standard in AI tools thanks to the CUDA platform, supported by PyTorch, Ollama, ComfyUI and others. We deliver AI workstations with Linux or Windows and current NVIDIA drivers.
Recommended configurations
Ultra and Ultra Intel with the RTX 5090 are a good choice for local language models and image generation. We recommend Threadripper for training and multi-GPU work. Systems with more cards or RTX PRO cards are quoted individually.
All configurations →Which language model can I run on an RTX 5090?
A card with 32 GB of VRAM can run models of around 30 billion parameters entirely on the GPU in quantised versions, and smaller models with plenty of room for context. Models of around 70 billion parameters need more memory, meaning several cards or an RTX PRO card.
Do two graphics cards add up their memory?
They do not form one shared memory, but popular tools such as llama.cpp and vLLM can split a model across several cards. This makes larger models possible, although some performance is lost to communication between the cards.
Windows or Linux for AI?
Most tools run on both. Linux usually makes deploying libraries and containers simpler, while Windows can be more convenient if you also use graphics applications on the same workstation. We install the system of the customer's choice.