Search Ctrl K
Home / Tools
THE FEUTEX TOOLKIT

GPU Kernel Launch Overhead

Estimate wall time attributable to launching many small kernels.

Developer workspace● Runs in your browser
01 Your input
02 Your result
View or copy raw result data
FormulaSequential elapsed time = n*(launch+work)/1000; Launch overhead share = launch+work===0?0:100*launch/(launch+work)

Made for real tasks

  • Estimate wall time attributable to launching many small kernels.
Explore step-by-step workflows →

Worked example

Inputs

Kernel launches
10000 launches
Unhidden launch overhead
5 microseconds
Device computation per kernel
10 microseconds

Expected results

Sequential elapsed time
150 ms
Launch overhead share
33.33333333 %

Load the example in the workspace to reproduce these results. Display rounding does not change the underlying calculation.

How to use this tool

  1. Enter kernel launches, unhidden launch overhead, device computation per kernel in the displayed units.
  2. Calculate the scenario, then compare it with measurements before changing a system.

Try this example

{
  "n": 10000,
  "launch": 5,
  "work": 10
}

Accuracy and scope

Sequential unhidden launch cost; asynchronous overlap, graphs and kernel fusion change behavior.

Is my data uploaded?

This tool runs locally in your browser. Input and output are not sent to a processing server. Copy or download your result before leaving the page.