Only plans with a discrete GPU are listed here: for running and training neural networks, working with LLMs, image generation, and rendering. The table shows the GPU model and its memory, while the “Server type” filter lets you display only dedicated servers or only VDS.
Exchange rates as of 29.10.2026 21:41 UTC (ECB, NBU, CBR, Bybit USDCUSDT). ≈ marks converted prices.
ishosting 6.1
HOSTKEY 8.6
Fornex 6.9
HOSTAFRICA 5.8
HOSTKEY 8.6
The model must fit entirely into the GPU memory (VRAM) along with its context. If there is not enough memory, the model either will not start or will run several times slower.
As a rough estimate for running a model, each billion parameters requires about 2 GB of memory in 16-bit format, about 1 GB in 8-bit format, and 0.5–0.6 GB in 4-bit (quantized) format, plus 10–20% for the context.
| GPU memory | What roughly fits |
|---|---|
| 8–12 GB | Models with up to 7–8 billion parameters in 4-bit format, Stable Diffusion, Whisper |
| 16–24 GB | Models with up to 13–14 billion parameters in 8-bit format or up to 30 billion in 4-bit format, SDXL, and Flux |
| 32–48 GB | Models with up to 30 billion parameters in 8-bit format or up to 70 billion in 4-bit format |
| 80 GB and more | 70-billion-parameter models in 8-bit format and fine-tuning large models |
Fine-tuning requires several times more memory than running the same model; LoRA significantly reduces these requirements.
A model can be split across multiple GPUs (vLLM, llama.cpp, and other tools support this), in which case their memory is combined. Data transfer between GPUs is slower than data transfer within a single GPU, so two 24 GB cards will run slower than one 48 GB card, but this is usually sufficient for running models. The number of cards is shown before the model: “2× …”, “4× …”; the memory listed in the table is for a single card.
