NVIDIA / nvbench

CUDA Kernel Benchmarking Library
Apache License 2.0
474 stars 63 forks source link

Add option for reporting min/max/median runtime #135

Open sleeepyjack opened 1 year ago

sleeepyjack commented 1 year ago

I stumbled across this Slack thread recently when I was trying to measure small kernels with nvbench and got fluctuating results. As @senior-zero notes in the thread, the variance of the minimum runtime is usually smaller than that of the average runtime.

So it would be very useful if nvbench would provide a way to generate these values directly. We could also report additional metrics, such as max or median runtime.