Profiling bank conflicts with Nsight Compute: a hands-on guide
anbarth · x · 2026-11-16
Antje Barth shares an actionable recipe for spotting shared-memory bank conflicts: profile your kernel — or your serving engine — with NVIDIA's profiler, Nsight Compute, with source counters enabled.
- The Source page shows each instruction's wavefronts (passes), the ideal count, and the excess
- Any excess on a shared-memory read indicates a bank conflict
- She cautions this measures a single read, not end-to-end tokens per second
This complements her earlier swizzling explainer on vLLM's quantized MoE kernel in the same thread — one diagnoses, the other fixes.
More from Infra
- GPU bank conflicts can slow reads 32x: why vLLM deliberately scrambles memory — anbarth · 2026-11-16
- Sentry CEO: Durable Objects complexity is Cloudflare platform's biggest weakness — zeeg · 2026-11-16
- WSJ: Nvidia's days as the AI economy's open checkbook are numbered — Gregory_C_Allen · 2026-11-16
- Survey maps how LLMs now deploy on edge devices via software optimization and hardware — kalyan_kpl · 2026-11-16
- SemiAnalysis layer-by-layer teardown reveals how Kirin 9050 implements LogicFolding — teortaxesTex · 2026-11-16
- Analyst: US chip export ban will backfire, forcing China to build own AI stack in 3-10 years — tengyanAI · 2026-11-16