A ocupação não é tudo

It is tempting to chase 100% occupancy, but maximum occupancy does not guarantee maximum performance.

Occupancy matters only because it helps hide latency. Once latency is hidden, extra warps add nothing useful.

Many kernels reach peak speed at just 50% occupancy. Beyond that point you hit diminishing returns and tuning elsewhere pays more.

A thread doing several independent operations hides latency on its own, so it needs fewer resident warps to stay busy.

Sometimes giving threads more registers lowers occupancy yet speeds the kernel, because it cuts memory traffic and spills.

If a kernel is limited by memory bandwidth, adding warps will not help. Better access patterns will.

When the math units are saturated, the kernel is compute bound. More occupancy cannot push past the arithmetic ceiling.

Let Nsight Compute tell you whether you are memory bound or compute bound before you touch the launch config.

The only reliable judge is the clock. Try a few block sizes and keep the one that runs fastest on real data.

Treat low occupancy as a clue, not a verdict. Investigate why it is low, then decide if raising it actually helps.

Good tuning balances occupancy, register use, and memory behavior together, optimizing the real bottleneck rather than one metric.

Recall when more occupancy stops helping a kernel.

You learned that occupancy is a means, not the goal: hide latency, find the real bottleneck, and let benchmarks decide. 🏁