MoE Routing Bottlenecks Identified in Edge Memory Bandwidth
August 20, 2026
A study of Qwen3-235B shows that serving large Mixture-of-Experts models on consumer hardware is limited by memory bandwidth rather than compute, resulting in 0.44 tok/s due to SSD paging thrash.
HOW THIS AFFECTS YOU
●
builderYou must account for the memory bandwidth wall when deploying large MoE models on edge hardware.
●
researcherThe study suggests that training for router locality could mitigate these hardware bottlenecks.