Mixing an RTX 5080 with a 3080Ti to offload CLIP/VAE for local video gen models
MonkeyCartridge · reddit · 2026-10-11
The author added an RTX 5080 but kept a 3080Ti in a secondary PCIe slot, aiming to run RAM-offload-heavy video gen models like Krea2 and Minimax H3 by loading CLIP and VAE onto the older card so the 5080's VRAM is reserved for the UNet.
MultiGPU node setups (e.g. buqi H3 multigpu) turned out clunky, sometimes requiring routing through WSL. He also weighs common alternatives — rendering on one GPU and upscaling on the other only reduces model swapping, while per-GPU workers don't pay off outside batch generation — and asks the community how to best leverage mixed multi-GPU rigs for larger models.
More from Infra
- Florida's 3 biggest utilities form alliance to pave way for data centers — 4KTV · 2026-10-11
- China has 40 nuclear reactors under construction, 20 more in final approval — LinusEkenstam · 2026-10-11
- TensorFold 1.0.7 writes each learned fact into ~10 new neurons, 4x faster with 3D view — HankYeomans · 2026-10-11
- China building 38 nuclear reactors (~39.8GW) as energy emerges as AI's biggest bottleneck — kimmonismus · 2026-10-11
- Token prices keep falling, yet devs burn more: Jevons paradox hits AI coding agents — daniel_mac8 · 2026-10-11
- Tenstorrent Blackhole folds 4x more per dollar than H200; antibody docking hits 84% top-1 — DavidBennett__ · 2026-10-11