Running AI on e-waste GPUs: 2.5 hours to render a 15s video on dual RTX 2060s
DevGnoll · reddit · 2026-11-14
A hobbyist on a pre-pandemic rig (dual RTX 2060 8GB + 128GB RAM) shares findings: ComfyUI only uses half the available VRAM, and a 15-second video with 3 image references takes 2.5 hours to render. On the LLM side, forked llama.cpp builds let him run solid Qwen Coder / Gemma models at faster-than-reading token speeds by saturating VRAM, 3/4 of system RAM, and 75% CPU.
He's asking whether similar "use all compute" abominations exist for ComfyUI.
More from Infra
- NVIDIA and TSMC reportedly never signed a formal contract in nearly 30 years of partnership — IgorCarron · 2026-11-14
- SemiAnalysis: SpaceX Memphis data centers to hit 2.4 GW by end of 2026, from 0.4 GW in January — mark_k · 2026-11-14
- SemiAnalysis warns VCs may lure away NVIDIA execs with billions as AI chip gold rush heats up — elonmusk · 2026-11-14
- Tesla Optimus plus Starlink could put robots in remote sites before full autonomy arrives — jeasinema · 2026-11-14
- Anyone actually running local LLMs on a DGX Station B300? Seeking real-world numbers — spaham · 2026-11-14
- Lawyer's 8x V100 private AI server runs 80% of DeepSeek V4.1 Flash at 4-bit nvfp4 — General-Spite1222 · 2026-11-14