vLLM Uses Swizzling to Eliminate GPU Shared Memory Bank Conflicts
GPU shared memory bank conflicts can slow warp reads by up to 32x; as Antje Barth demonstrates with vLLM v0.33.0's quantized MoE kernel, a one-line XOR-based swizzle eliminates them.
2026-11-16 ~ 2026-11-16 · 3 related posts
- GPU bank conflicts can slow reads 32x: why vLLM deliberately scrambles memory — anbarth · 2026-11-16
- GPU bank conflicts can slow reads 32x: why vLLM deliberately scrambles memory — anbarth · 2026-11-16
- How vLLM v0.33.0 swizzles quantized MoE kernels to kill bank conflicts — anbarth · 2026-11-16