Guide4 sections2 min read
The short answer
For most Stable Diffusion and Flux work, a 24 GB or 32 GB GPU such as the RTX 4090 or RTX 5090 is the usual starting point. Choose 48 GB (L40S) or 96 GB (RTX PRO 6000) for larger models in full precision, higher resolutions, video models or large batches.
Memory decides what runs
Image models load their weights, text encoders and VAE into GPU memory. Check the parameter count on the model card and apply the same rule as for language models: about 2 bytes per parameter in FP16.
Resolution, batch size, ControlNets and LoRAs add memory. If a model does not fit, quantized or offloaded versions trade speed for memory.
Speed and cost per image
Once the model fits, newer architectures and higher bandwidth generate faster. Compare cost per image, not hourly rate: a faster GPU at a higher price can be cheaper per result.
Fine-tuning a LoRA usually fits on 24 GB; full fine-tuning of large image models needs more memory.
Pay the price you choose
If the best offer is above your budget, place a bid at the price you are willing to pay per GPU-hour. The bid is matched when an offer meets it.
During pre-launch, prices shown on EverXCloud are indicative and become rentable on launch day. The table above is read from the live exchange each time the page is generated.
Live prices on EverXCloud
| GPU | Memory | From, per GPU-hour |
|---|---|---|
| NVIDIA GeForce RTX 4090 24GB | 24 GB | $0.33 |
| NVIDIA GeForce RTX 5090 32GB | 32 GB | $0.69 |
| NVIDIA L40S 48GB | 48 GB | $0.79 |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | $1.69 |
Best live offer on the exchange, excluding tax. Prices are indicative during pre-launch.
Frequently asked questions
- Is 24 GB enough for Flux?
- It depends on the Flux variant and precision. Check the parameter count on the model card, compute about 2 bytes per parameter in FP16, and use a quantized version or a larger GPU if it does not fit.
- Which GPU is best for SDXL?
- A 24 GB GPU such as the RTX 4090 runs SDXL comfortably at standard resolutions. Larger batches or video models benefit from 48 GB or more.
- Are the prices shown final?
- No. During pre-launch, EverXCloud prices are indicative. They are shown per GPU-hour, excluding tax, and become rentable on launch day.