Server with RTX 5090: 32 GB of GDDR7 and what to watch with drivers and CUDA
The GeForce RTX 5090 is the top Blackwell gaming card, released in early 2025. According to the NVIDIA specification, it has 32 GB of GDDR7 on a 512-bit bus (about 1.8 TB/s), 21,760 CUDA cores, fifth-generation Tensor Cores with FP4 support and a power draw of up to 575 W. There is no NVLink. Compared with the RTX 4090, it has a third more memory, and that memory has about 78% more bandwidth.
What it is good for
- Language models up to 30–32B parameters in 4- to 8-bit quantization with room for context. Generation speed is limited by memory bandwidth, and here the RTX 5090 outpaces many data-centre cards.
- Image and video generation: 32 GB lets you keep large models fully in memory without offloading weights.
- Rendering and video encoding: three hardware NVENC encoders with AV1 support.
What to check in a plan
- CUDA version. Blackwell cards need CUDA 12.8 or later and recent framework builds (PyTorch 2.7 and up). On an old image a library may fail to detect the card even though nvidia-smi sees it.
- Power and cooling. 575 W per card calls for a strong power supply and good heat dissipation. For a dedicated server, ask what PSU the machine has, especially for two-card builds.
- CPU, RAM and disk. For inference, 8–16 cores, 64–128 GB of RAM and NVMe of 1 TB or more are enough.
- Plan type. Most offers are dedicated servers; cloud plans with this card are few.
Monthly and hourly cost
In the catalogue as of 17 November 2026, the median price of a server with one RTX 5090 is about $700 a month, with half of the plans between $650 and $830. Over 730 hours that comes to about $0.96 an hour. Fewer than half of the hosts with this card offer hourly billing. If 32 GB is not enough for your model, compare prices for the 48 GB cards, L40S and RTX A6000, or for the RTX PRO 6000 with 96 GB. More on choosing a card in which graphics card to choose for a server.