Under ten seconds per SDXL image. That threshold matters because it means you can generate, inspect, and tweak faster than you can write a new prompt. The RTX 3090 hits 9.38 seconds per 1024×1024 image on our GigaGPU dedicated servers, with 17.5 GB of VRAM headroom for advanced pipelines.
Speed Benchmarks
| Metric | Value |
|---|---|
| Iterations/sec | 3.2 it/s |
| Seconds per image | 9.38 sec (30 steps) |
| Images per minute | 6.4 |
| Resolution | 1024×1024 |
| Sampler | Euler a / DPM++ 2M Karras |
| Performance rating | Very Good |
30-step, 1024×1024, FP16, batch size 1. The 3090’s 936 GB/s memory bandwidth is a substantial advantage for diffusion workloads, which are bottlenecked by how fast weights and latents can be shuffled through the denoising loop.
VRAM Surplus
| Component | VRAM |
|---|---|
| Model weights | 6.5 GB |
| Sampling buffer | ~1.3 GB |
| Total RTX 3090 VRAM | 24 GB |
| Free headroom | ~17.5 GB |
Seventeen and a half gigabytes free after SDXL loads. That opens up serious pipeline possibilities: run the base model and refiner simultaneously, stack half a dozen LoRAs, generate at 2048×2048 natively, or use batch sizes of 4-8 for throughput-oriented workflows. You could even run SDXL alongside a lightweight LLM for prompt engineering automation.
Generation Costs
| Cost Metric | Value |
|---|---|
| Server cost | £0.75/hr (£149/mo) |
| Cost per 1K images | £1.95 |
| Images per £1 | 513 |
At £1.95 per thousand images, the 3090 is actually the most cost-effective SDXL card in the lineup on a per-image basis. It costs less per image than the 4060 (£2.08) or 4060 Ti (£2.19), while generating images more than twice as fast. The higher monthly cost (£149) is offset by the sheer volume you can push through. Calculate your own scenario with the benchmark comparison tool.
Who This Is For
Production image generation at scale. E-commerce catalogues, marketing asset pipelines, print-on-demand businesses, and any workflow generating hundreds or thousands of images per day. The 3090’s combination of speed, VRAM depth, and per-image cost makes it the workhorse card for SDXL.
Deploy:
docker run --gpus all -p 7860:7860 ghcr.io/ai-dock/stable-diffusion-webui:latestMore: SDXL hosting guide, best GPU for Stable Diffusion, benchmark archive. Related: Flux.1 hosting.
SDXL at 6.4 Images/Min — RTX 3090 Servers
The best per-image cost in the lineup. UK datacentre, £149/mo, root access.
Provision RTX 3090
