Best Cloud GPU for Stable Diffusion
For most Stable Diffusion work the value pick is an RTX 4090 at $0.69/hr: 24GB of VRAM runs SDXL comfortably and its raw image throughput beats far pricier datacenter cards. If you batch heavily or serve an endpoint, an L40S trades a little speed for 48GB and datacenter reliability. Reach for an A100 only when you are training or fine-tuning a model rather than just generating images.
The picks, with live prices
| Pick | GPU | VRAM | On-demand from | Where | |
|---|---|---|---|---|---|
| value pick | RTX 4090 | 24 GB | $0.69 | RunPod secure cloud | Rent → |
| performance pick | RTX 5090 | 32 GB | $0.86 | Spheron on-demand | Rent → |
| scale pick | L40S | 48 GB | $0.88 | Massed Compute on-demand | Rent → |
| budget pick for training | A100 | 80 GB | $0.89 | Jarvislabs on-demand | Rent → |
RTX 4090 value pick
24GB VRAM covers SD 1.5 and SDXL including higher resolutions and modest batches. Consumer Ada silicon gives the best image-per-dollar of any card here, so single-user generation and LoRA training both land here first.
RTX 5090 performance pick
32GB and a newer architecture push more iterations per second than a 4090, which shortens big batch and high-step runs. Pick it when generation time is the bottleneck and the hourly rate still pencils out against the extra speed.
L40S scale pick
48GB and datacenter-grade uptime suit an inference endpoint that keeps multiple pipelines or a large ControlNet stack resident. Slower per image than a 4090 but far more headroom, and it is available on more clouds for production.
A100 budget pick for training
Overkill for plain image generation, but the 80GB variant is the sensible floor for full fine-tunes or training a model from a checkpoint. If you are only sampling images you are paying for memory bandwidth you will not use.
Worth knowing
- SD 1.5 fits in 8 to 12GB; SDXL is comfortable at 16 to 24GB once you add refiner, ControlNet, or upscaling.
- Generation is compute-bound, not memory-bound, so a fast consumer card usually beats a datacenter card that costs several times more per hour.
- VRAM sets your ceiling on batch size and resolution, not your speed. Buy memory for bigger batches, buy clock speed for shorter runs.
- For a hosted endpoint, per-second billing and fast cold starts matter more than the sticker rate, since idle time dominates cost on bursty traffic.
FAQ
SD 1.5 runs in 8 to 12GB. SDXL wants 16 to 24GB once you stack a refiner, ControlNet, or high-resolution upscaling. A 24GB card like the RTX 4090 handles nearly everything short of full model training.
Not for plain sampling. Image generation is compute-bound, so a consumer RTX 4090 or 5090 produces more images per dollar. The A100's large VRAM only pays off when you fine-tune or train, not when you generate.
Prices render from today's verified snapshot, not from when this guide was written. Full table on the homepage; break-even math in the calculator.