LLM inference
Serve models that fit in 32 GB, often quantized, with vLLM, SGLang or Ollama.
Rent NVIDIA RTX 5090 GPUs with 32 GB VRAM
RTX 5090 servers with 32 GB of GDDR7 memory on the Blackwell architecture, for LLM inference, image generation and CUDA development. Check current configurations and prices, then create your server from the console.
- 32 GB GDDR7 per GPU
- Templates for CUDA 12.8 and 13.0
- Single and multi-GPU configurations when listed
ONIE rents NVIDIA RTX 5090 GPU servers with 32 GB of GDDR7 VRAM from a live inventory. Pick a template built for CUDA 12.8 or newer, such as PyTorch, vLLM or ComfyUI, choose a disk size and pay hourly or by billing cycle; then connect over SSH or the browser terminal.
Available GPUs and current prices
Live inventory: availability and prices change with the market. The console confirms the final price before you order.
Loading current prices…
RTX 5090 at a glance
- GPU
- NVIDIA GeForce RTX 5090
- Memory
- 32 GB GDDR7 per GPU
- Architecture
- Blackwell
- Software
- CUDA 12.8 or newer (templates with CUDA 13.0 available)
What RTX 5090 servers are used for
Image and video generation
ComfyUI workflows with Flux, SDXL and video models.
Fine-tuning
LoRA-style fine-tuning of smaller models with Unsloth Studio.
CUDA development
Ubuntu 24.04 with CUDA 13.0 for building and testing GPU code.
Check the minimum VRAM of a template before ordering; models larger than 32 GB need quantization or a multi-GPU configuration.
How to rent a GPU
- 01
Choose a template and disk size
- 02
Pick a compatible GPU from the live inventory
- 03
Confirm the price and create the server from your account balance
- 04
Connect over SSH or the browser terminal
How GPU billing works
- Hourly: the GPU is charged per hour while the server runs. A stopped hourly server is only charged for its disk.
- Billing cycles: monthly, quarterly, semi-annual or yearly, with any discount shown before you pay.
- Disk storage is charged per GB together with the GPU price.
- Prices are in USD; you can display them in VND at the current exchange rate.
Frequently asked questions
How much VRAM does the RTX 5090 have?
32 GB of GDDR7 per GPU. In a multi-GPU configuration, each GPU has its own 32 GB.
Which CUDA version do I need?
The RTX 5090 uses the Blackwell architecture and needs CUDA 12.8 or newer. The RTX 5090 templates use CUDA 12.8 or 13.0, and only compatible servers are offered.
How is an RTX 5090 server billed?
Hourly while it runs, or by monthly, quarterly, semi-annual or yearly cycle. Disk is charged per GB, and a stopped hourly server is only billed for its disk.
Can I run large language models on it?
Yes, models that fit in 32 GB of VRAM, often with quantization, or larger models across several GPUs when a multi-GPU configuration is available.
What if no RTX 5090 is available?
Availability follows the live inventory. Check back later, or choose another GPU with enough VRAM on the GPU Cloud page.