Why Flux slows down on an 8GB card: ComfyUI silently spills to CPU when VRAM fills
leonbuilds · reddit · 2026-08-31
After three weeks of measurement on a 5060 8GB under WSL2 running Flux and Wan, the author found every slowdown came not from GPU compute but from VRAM filling up and ComfyUI silently offloading work to CPU/disk — no error, no UI warning, only a 0.00 MB usable line in the log. Worst case: adding an upscaler to a Flux fill outpaint graph pushed per-image time from 97s to 240s.
Takeaways: watch the card, not the clock — a healthy run sits at 100% util, 34-56W, 63-77°C with VRAM nearly full and holding; anything less suggests spilling. One caveat: model loading looks identical (5% util, 6.2W) and is fine — read the log before killing anything. The author also asks how to make ComfyUI warn on the 0.00 MB usable case.
More from Infra
- Europe invests €387.8M in LUMI-AI supercomputer with 10x AI capacity — wkmyrhang · 2026-09-01
- SK hynix reportedly considers Intel Foundry for HBM4E base dies — AccBalanced · 2026-09-01
- Guardian proposes off-grid, self-powered datacenters to cut emissions — nordicinst · 2026-09-01
- Compute Wants to Leave Earth: A Manifesto for Orbital Infrastructure — McDonaghMatthew · 2026-09-01
- Optimizing Qwen 3.8 Flash Next: Improving speeds on 64GB VRAM setup — Jorlen · 2026-09-01
- How much power does the AI buildout actually take? — TheZachMueller · 2026-08-31