Table of Contents
SDXL is the most forgiving major image model on hardware — fits comfortably on 8 GB cards, works with offloading on smaller. But for production image-generation APIs with ControlNets, LoRAs and refiners stacked, the right GPU still matters.
For SDXL: RTX 5060 Ti 16 GB is the cost leader at £119/mo with ~7s per 1024² image. RTX 5090 is the throughput leader at ~2.5s per image with comfortable headroom for refiner ensembles. Below 16 GB VRAM, ControlNet stacks pinch.
GPU ranking
- RTX 5090 32 GB (£399/mo) — fastest. Hosts SDXL + refiner + multiple LoRAs comfortably.
- RTX 5080 16 GB (£189/mo) — fast, FP8 path, tight on full ensemble.
- RTX 5060 Ti 16 GB (£119/mo) — best cost-per-image. Sweet spot for moderate volume.
- RTX 4090 24 GB (£289/mo) — Ada hardware. Slightly slower than 5080, more VRAM.
- RTX 3090 24 GB (£159/mo) — older but 24 GB matters for big ControlNet stacks.
- RTX 3060 12 GB (£99/mo) — cheapest credible. Tight but functional.
By workflow
- SDXL Base only → 5060 Ti 16 GB
- SDXL + ControlNet → 5060 Ti or 5080
- SDXL + Refiner ensemble → 5080 or 5090
- SDXL + multiple LoRAs hot-loaded → 5090 (24 GB+ recommended)
- SDXL Turbo + batch generation → 5090 (highest throughput)
- SDXL + SDXL Refiner + Inpaint + Upscale pipeline → 5090 or 6000 Pro
Verdict
SDXL is the rare workload where the cheap GPU is genuinely good enough for production. RTX 5060 Ti 16 GB hosts a real SDXL API at £119/mo. Step up only when ControlNets or LoRAs stack into the VRAM ceiling.
Bottom line
For SDXL specifically the cheapest viable card is the right one until you hit a ControlNet wall. For deeper sizing see SDXL VRAM requirements and RTX 5090 + SD verdict.
