Question Why isn't --no-graphics faster than with graphics?

Hey there,

I started training without graphics (setting --no-graphics in mlagents-learn command) and would have imagined it to be quite a bit faster than training without graphics - but it actually performs the same as with graphics?

Somehow this is counterintuitive. Do you guys know why this is the case?

 

If it's a CPU bottleneck speed doesn't change if the GPU is not used or half used. Power usage could be improved tho

 

I noticed that too. Kind of curious too...

One thing I did notice is that by default the Update rate is hundreds of frames per second, which chews up all available CPU. I don't think that does much to help the training time.

I usually drop the Update rate to 10fps, e.g. https://github.com/hughperkins/peac...cf8ec3b6db48d49/PeacefulPie/Simulation.cs#L87

Code (csharp):

1.  
2. Application.targetFrameRate = 10;
3.  

I use the following method to detect whether I'm running in dedicated mode https://github.com/hughperkins/peac...c3b6db48d49/PeacefulPie/Simulation.cs#L61-L63 :

Code (csharp):

1.  
2. public bool isDedicated() {
3.     return Screen.currentResolution.refreshRate == 0;
4. }
5.  

Not sure if dropping the frame rate a bunch changes your results? It should at least render the question relatively mute, since the graphics will no longer take up much time anyway?