Quick Answer

For local AI and Stable Diffusion in Paarl, VRAM is the deciding spec: an RTX 4060 Ti 16 GB runs SDXL at roughly 1.5-2.5 iterations/sec and is the value entry point around R10,500. For serious throughput, an RTX 4070 Ti Super 16 GB or RTX 4090 24 GB (the latter clears SDXL in under 4 seconds per 1024px image) is worth the jump. CPU and RAM matter less than GPU VRAM here.

Why VRAM, not raw speed, decides your Stable Diffusion rig

Stable Diffusion and local image models load the entire model into GPU memory. An 8 GB card (RTX 4060) handles SD 1.5 comfortably but chokes on SDXL and Flux at higher resolutions. The practical floor for SDXL and ComfyUI workflows is 12 GB; 16 GB lets you run Flux.1 and larger LoRAs without out-of-memory errors.

Recommended tiers currently stocked at Evetech:

  • Entry (SD 1.5, light SDXL): RTX 4060 8 GB, ~1.2 it/s SDXL 512px
  • Value (SDXL, Flux): RTX 4060 Ti 16 GB, ~2 it/s SDXL 1024px
  • Performance: RTX 4070 Ti Super 16 GB, ~4-5 it/s
  • Pro: RTX 4090 24 GB, fastest single-card option

Supporting components

Pair the GPU with 32 GB DDR5 and a fast NVMe SSD (1 TB minimum) because model checkpoints are 2-7 GB each and you will collect many. A Ryzen 5 7600 or Core i5-13400F is plenty; the CPU only feeds data to the GPU. A quality 650-750 W 80+ Gold PSU keeps a 4070 Ti Super stable.

FAQ

How much VRAM do I need for Stable Diffusion?

8 GB runs SD 1.5, but SDXL and Flux need 12-16 GB to avoid out-of-memory errors at 1024px and above. For local AI as a serious hobby, a 16 GB card like the RTX 4060 Ti 16 GB is the practical minimum.

Can I run local LLMs on the same rig?

Yes. The same 16 GB+ VRAM card runs quantised 7B-13B LLMs locally. For 70B models you need 24 GB (RTX 4090) or multi-GPU, so size the GPU to your largest planned model.

Will Evetech deliver to Paarl?

Yes, Evetech ships to Paarl and the wider Cape Winelands in 2-3 business days. Power-test the GPU on arrival and confirm the model boots before signing for it.

TIP

or Automatic1111, enable xFormers and use fp16 to roughly halve VRAM use, letting a 16 GB card handle batch sizes that would otherwise need 24 GB.