clbr / radeontop

GNU General Public License v3.0
789 stars 69 forks source link

Feature Request: Box/Violin Plots #92

Open Bengt opened 4 years ago

Bengt commented 4 years ago

While RadeonTop draws 120 samples per second, it only displays the mean of the utilisation levels. The mean already informs the user about which components of the GPU are being utilised at all, but it does not provide the information necessary for determining which block of the GPU is limiting performance.

For example, I am training neural networks and am thus interested in whether my training process is starved for memory bandwidth, clock speed or something else entirely. The means show that memory and clock speeds are not constantly running at their maximum frequencies, but since they both clock up and down multiple times per second, I cannot determine which one limits first, thus giving me no information on how to tune my training process next.

A more detailed visualisation of the samples would be to be box and whisker plots or violin plots like these:

https://github.com/ewheeler/ascii-boxplot

Money quote:

test data      ---------------==========O==========---------------
trek00 commented 4 years ago

if you need to diagnose a timing issue, may be this commit trek00/radeontop@a228b31 can be useful; try with a command line like: radeontop -t 100000 -i 0.1