search

nihui / ncnn-small-board

ncnn benchmark on various single board computers
158 stars 7 forks source link

The running time with vulkan in jetson nano is much slower than yours #2

Open twmht opened 2 years ago

twmht commented 2 years ago

Trying to run benchncnn in jetson nano (Jetpack 4.6) with vulkan.

ncnn version is latest.

But the running time is much slower than yours

Here are my result

image

Any idea?

nihui commented 2 years ago

the data is collected from https://github.com/Tencent/ncnn/issues/3485

https://qengineering.eu/install-ncnn-on-jetson-nano.html Note, during the test the Jetson Nano CPU was overclocked to 2014.5 MHz and the GPU to 998.4 MHz.

Qengineering commented 2 years ago

There are several options to maximize the performance of the benchmark. There are several options to maximize the performance of the benchmark. First, wait 10 minutes after boot up. It gives Ubuntu (Tegra) the time to do all the administrative tasks like looking for updates etc. Second, set the power to 10 Watts, something you've probably already done. Thirdly, set all the governors at performance.

$ sudo echo "performance" > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
$ sudo echo "performance" > /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
$ sudo echo "performance" > /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor
$ sudo echo "performance" > /sys/devices/system/cpu/cpu3/cpufreq/scaling_governor

You can temporarily shut down the graphic stack or start overclocking. Both methods are not used here. The results with a Jetson Nano 4 GB running at a normal clock speed of 1479 MHz CPU and 921.6 MHz GPU.

BenchNCNN