12 November 2026 · VRAMfarm team

RTX 5090 for AI: what 32 GB of GDDR7 runs that a 24 GB card can't

Eight gigabytes sounds like a small step, but on the RTX 5090 it is the difference between “fits with tricks” and “fits”: a 32B language model with 30–40k tokens of context instead of 8–10k, Wan 2.2 14B video in fp8 at 720p, and Flux in full bf16. The card also reads memory 78% faster than an RTX 4090 (1,792 vs 1,008 GB/s), so it generates text about 1.4–1.7 times faster and images about 1.3–1.5 times faster. It rents from €0.76 an hour, against €0.59 for the 4090. The one catch is software: Blackwell needs CUDA 12.8 or newer.

What changed from the RTX 4090

The pattern is the opposite of the step from the RTX 3090 to the 4090 (see RTX 3090 vs RTX 4090). That step doubled compute and barely moved bandwidth. This one adds about 27% compute and 78% bandwidth, plus a third more memory. So the 5090 gains most on language models, which are limited by bandwidth, and less on image models, which are limited by compute.

Models that fit now

These are estimates for typical settings; context length, the runtime’s own buffers and ComfyUI’s offloading options move them by a few gigabytes. The VRAM guide has the formula behind them.

The language model rows matter most in daily use. On a 24 GB card, a 4-bit 32B model leaves 2–3 GB for the KV cache, which is why long documents end in “CUDA out of memory”. On 32 GB the same model has roughly four times as much room. Alternatively, spend the extra memory on a better quantisation: a 5- or 6-bit 32B model is a little closer to the original than the 4-bit one, especially for code and maths.

For video, Wan 2.2 14B is the headline. A 5-second 720p clip in fp8 needs about 30 GB, so the 5090 is the smallest card that runs it without offloading. Longer clips, higher resolution or extra nodes push past 32 GB, which is why our ready-made Wan 2.2 14B app asks for a 48 GB card. On a 5090 you can still run it yourself in an empty ComfyUI and keep clips short.

How much faster than a 4090, and what a job costs

Image and video times assume a ComfyUI queue that keeps the card busy; the token figures are for one user in Ollama or llama.cpp. They depend on the workflow, sampler, prompt length and software version, so treat the ratios as the useful part.

The 5090 is cheaper per job whenever its hourly price is below its speed advantage times the 4090’s price: below about 1.3–1.5 times for images and training, below about 1.4–1.7 times for chat. Compare the two prices on the pricing page when you launch; catalogue prices are live, and the price per minute is fixed once your server starts. For work in front of the screen, add your own time: a Flux image every 7–10 seconds instead of every 12 changes how a session feels.

CUDA 12.8 or newer: the one catch

Blackwell cards use a new architecture, sm_120, and older software has no code for it. To run anything on a 5090 you need:

• an NVIDIA driver from the 570 series or newer on the machine (580 or newer if you use a CUDA 13 build);
• a PyTorch build for CUDA 12.8 or later (cu128, cu129 or cu130);
• PyTorch 2.7 or newer, the first stable release with Blackwell support.

A wrong build fails with “no kernel image is available for execution on the device”, often after a warning that the card “with CUDA capability sm_120 is not compatible with the current PyTorch installation”. The same happens with extensions pinned to old versions: xformers, flash-attn, bitsandbytes, custom attention kernels or ComfyUI nodes that reinstall an old torch.

How to avoid it on a rented machine:

1. Check the machine before launch. The catalogue shows each machine’s driver and CUDA version, and the “CUDA, from” filter hides older ones. Any machine with a 5090 should already have a recent driver, but a filter at 12.8 costs nothing.
2. Use a current stack. Our ready-made apps and the empty ComfyUI come with PyTorch for CUDA 12.8. In your own environment, install with pip install torch --index-url https://download.pytorch.org/whl/cu128.
3. Keep old projects on older cards. A repository pinned to torch 2.1 and CUDA 11.8 will not run on Blackwell at all. Run it on a 4090 or 3090, or update its dependencies.

The CUDA and driver mismatch guide has the full table of drivers, builds and error messages.

When you still need a 48 GB card

32 GB closes many gaps, but not all of them. Take a 48 GB card for:

• 70B language models. In 4-bit the weights alone are about 43 GB. A 48 GB card runs them with 8–12k tokens of context; an 80 GB card with long context.
• Longer or sharper Wan 2.2 14B video. More than 5 seconds at 720p, 1080p, or the model in bf16.
• 32B models in fp8 (about 33 GB), or serving 32B to several users with vLLM. Each parallel conversation needs its own KV cache.
• Qwen-Image in bf16 (about 41 GB for the image model alone) or heavy multi-LoRA image workflows.
• QLoRA on 32B models. 24–32 GB works with short sequences; 48 GB is comfortable.

The 48 GB options are the RTX A6000 from €0.72 an hour and the L40S from €1.06; the RTX PRO 6000 is a Blackwell workstation card with 96 GB from €2.12. The 48 GB cards have less bandwidth than a 5090 (768–864 GB/s), so they are slower at things that fit on both. Pick them for memory, not speed.

For full fine-tunes and 70B models in fp8, the next step is an 80 GB A100 or H100.

Who should rent the 5090

Rent it if you chat with or serve 14–32B models and want long context, if you make Wan 2.2 video, or if you run Flux, Qwen-Image or HiDream all day and want fewer model swaps. Stay on a 4090 if your workflow already fits 24 GB comfortably and the price ratio does not favour the 5090 that day, or if your project is pinned to an old CUDA stack. Stay on a 3090 for cheap long batches where nobody waits.

If you are unsure, the cheapest test is the real one: run the same short job on a 4090 and a 5090 and compare the minutes. With per-minute billing, a 15-minute test costs €0.15 on the 4090 and €0.19 on the 5090. The RTX 5090 page lists the machines available now, with their country, driver and CUDA version.