Profile, Optimize, Ship
Closing the loop with Nsight metrics.
Profile, Optimize, Ship is a free CUDA Academy lesson on CoddyKit — lesson 4 of 4. You can read the complete lesson below for free — then practise it hands-on in the browser with a built-in code editor and a 24/7 AI tutor. It is part of the CUDA Academy learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.
Measure, Do Not Guess
Optimization starts with data, not hunches. Always profile first to learn where the time really goes before you change a single line. 🔍
Start With the Timeline
Open Nsight Systems to see the whole run: kernels, copies, and gaps. The timeline shows whether transfers and compute actually overlap.
Find the Hot Kernel
One stage usually dominates. The timeline reveals the longest-running kernel, and that is exactly where your tuning effort should go first.
Zoom In With Nsight Compute
For the hot kernel, switch to Nsight Compute. It reports per-kernel metrics like memory throughput, stalls, and achieved occupancy.
Read the Roofline
The roofline tells you if a kernel is memory-bound or compute-bound. That single answer decides which optimizations are even worth trying.
Fix Memory-Bound Kernels
If you are memory-bound, chase coalescing, shared-memory tiling, and vectorized loads. Feeding data faster is the only path to more speed.
Fix Compute-Bound Kernels
If you are compute-bound, raise ILP with unrolling, cut redundant math, or move to lower precision. Here the arithmetic units are the limit.
Label Your Code With NVTX
Wrap pipeline stages in NVTX ranges so the timeline shows named bars instead of anonymous kernels. Readable profiles are faster to debug.
nvtxRangePush("blur");
blurKernel<<<grid, block>>>(d_in, d_out);
nvtxRangePop();Change One Thing at a Time
Apply a single optimization, then re-profile. This tight loop proves each change helped and stops you from chasing two effects at once.
Know When to Stop
When a kernel nears its roofline limit, more tuning yields little. Recognize diminishing returns and spend your time on the next bottleneck.
Validate, Then Ship
Before release, confirm correctness against a reference and check the speedup is stable. A fast but wrong pipeline is not shippable. 🚢
Quick Check
Nsight Compute says your hot kernel is memory-bound. Which fix should you try first?
Recap
You learned to profile with Nsight, read the roofline to classify kernels, apply targeted fixes one at a time, and validate before shipping a fast, correct pipeline. 🏁
Frequently asked questions
Is the “Profile, Optimize, Ship” lesson free?
Yes — the full text of “Profile, Optimize, Ship” is free to read here on the web, and the CUDA Academy course includes 4 lessons in total. To practise it interactively (a built-in code editor and a 24/7 AI tutor) and unlock the rest of the CUDA Academy course, upgrade to CoddyKit PRO.
What will I learn in “Profile, Optimize, Ship”?
Closing the loop with Nsight metrics. You practise CUDA Academy with hands-on code you run directly in the browser, and a 24/7 AI tutor answers your questions as you work through the lesson.
Do I need any experience to start CUDA Academy?
No prior experience is required. CUDA Academy on CoddyKit is structured for beginners through advanced learners; this is — lesson 4 of 4, so you can start here or from the beginning and move at your own pace.
How long does the “Profile, Optimize, Ship” lesson take?
Most CoddyKit lessons take about 5–10 minutes. Each one is bite-sized and interactive, so you make steady progress and pick up exactly where you left off across the web and the app.
Can I write and run code in this CUDA Academy lesson?
Yes. Every CUDA Academy lesson includes a built-in code editor, so you write and run real code right in your browser and get instant AI feedback — no local setup required.