Why two 5060 Ti GPUs still top out at 50% on a split 27B model

dsdt · reddit · 2026-07-22

The poster explains why two 5060 Ti cards running Qwen 3.6 27B only show about 50% utilization.
The bottleneck is memory bandwidth: generating a token requires reading the model weights, so the setup is limited by bandwidth rather than compute.
With the model split layer-by-layer across two cards, one GPU often waits while the other works, which makes 50% usage a ceiling rather than a bug.
They also note that row-split could use both cards on the same layer, but without NVLink the result depends heavily on PCIe lanes.
Google’s DiffusionGemma is mentioned as a different tradeoff: it generates 256 tokens in blocks to sidestep per-token overhead, but quality drops.

More from Infra