Effects of PCIe bandwidth on inference
retained text
So i have a new card in an old system. RX 9060 XT supports PCIe Gen5 x16 but due to 2 bottlenecks, my Motherboard supporting only Gen3, and my AMD APU limiting it to x8, the card runs on Gen3 x8. How does that affect inference of models, in both cases, fully in vram or offloaded partially to ram? At today's prices what I could replace it with is PCIe Gen4 x16. But would that be a meaningful upgrade?
document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xtioaa/effects_of_pcie_bandwidth_on_inference