GPU bank conflicts can slow reads 32x: why vLLM deliberately scrambles memory

anbarth · x · 2026-11-16

A bank conflict can make a GPU shared-memory read take 32 passes where 1 would do — and correctness tests never catch it. The author illustrates the mechanics (32 banks, one 4-byte value per pass per bank) and explains swizzling, the trick vLLM kernels use to deliberately store data scrambled and avoid conflicts.

Related event: vLLM Uses Swizzling to Eliminate GPU Shared Memory Bank Conflicts(3 posts)→

Original post →

More from Infra

Infra channel →