Profiling inference and when to write a kernel
Find where GPU time really goes with Nsight Systems, the PyTorch profiler and Nsight Compute; read a timeline, spot launch gaps and memory-bound kernels, and decide between flags, compiler fusion, a Triton kernel and a CUDA kernel.
17 min·Verified 2026-11Hands-on labPro+100 XP
Share
Everything so far in this course says what should make inference fast. A profiler says what is happening on your system. Teams that skip it spend weeks on the wrong optimisation. This lesson covers the tools, what to look for, and a sensible ladder from "flip a flag" up to "write a kernel".
Pro content
Keep reading with Inference Academy Pro
Unlock this full lab, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.