I see this tutorial presented by a bunch of different people https://randomnerdtutorials.com/esp32-deep-sleep-arduino-ide-wake-up-sources/

Here you can see they run a subroutine print_wakeup_reason() fairly early in the setup.  I have the same subroutine in my code, but I rarely ever see this printed (sometimes if I open a new terminal window at just the right time I can catch it).  Come to think of it, in this video https://youtu.be/919hSPthnb4 the gentleman uses "log_W" instead of "Serial.println".  He seems to be using a different terminal app all together or is that just the arduino IDE opened in a different window?  At any rate, his monitor seems to be catching everything. 

Hmmmm.....

As far as your original problem, I have no idea how one time it gives one value and another time a different value.

My approach would be to print the involved variables on the serial monitor to see if any of them look other than expected. I am talking about millis, fpstime, FPS, fpscounter. Set the serial monitor to display it's timestamp on every message as well, then you can use those timestamps to calculate what your millis should produce etc etc etc.

Good luck

What happens if instead of unplugging, you just press the reset switch ion the board?

    I have been trying to investigate deep sleep and it's use on esp32-C3.  One thing I noticed even before looking at this is the serial monitor on my setup seems problematic.  I am using platformIO but haver verified the same issues occur on arduino IDE.  One of the first issues I noticed was many pf the Serial.print statements in my setup() routine do not show up in the serial monitor.  It is almost like by the time the serial monitor connects, the code is through the setup already.  I do have the baud rate set to 115200 which is normal for esp32.  I have also changed it to other values like 9600 but no good.   A lot of my Serial.print statements are "missed" in the setup() (although occasionally it catches some of them).  Now when I started to investigate deep sleep the problem got even worse.  When my board goes into deep sleep, upon waking the serial monitor does not automatically re-establish.  I have to manually close the terminal then open a new one.  As you can guess, I miss a lot of data which is supposed to go to the serial port this way.  This makes it frustrating since I have no way to know what the behavior of my code is during a wake-up because  by the time I get a new terminal up, I am well into loop().

I have on-board JTAG debugging and that is also hampered since once the esp32 goes to sleep, the debugger gets lost and never re-connects.  

So how does this relate to my original issues of the FPS?  I did leave out a line of code in the loop() that was a Serial.println call which was the only other line of code in the loop.  I didn't think this would make a difference with my FPS.  However, because of my issues, once the code re-awakens, that Serial.println function somehow drags the code execution speed waaay down (if I comment out this line, FPS is 31 upon waking).  I think it is related to the other issues with my seial monitor.

Anyway, still trying to figure it out.

Fish

Where's the rest of the sketch ?

You haven't defined or declared u8g2.

You haven't defined or declared fpsTime (has to be unsigned long).

When you unplug USB and plug it back in milliseconds() will restart from 0.

After unplugging, how long does it take to get back to a reasonable FPS ?

int FPS;
int fpsCounter;
void printScreen(){
  u8g2.setCursor(110,64);
  u8g2.print(FPS);
  u8g2.sendBuffer();
  // get FPS
   if ((millis()-fpsTime)<=1000)
    { fpsCounter++;
    }
   else
   {  fpsTime=millis();
      FPS=fpsCounter;
      fpsCounter=0;
   } 
}
void setup() {
  Serial.begin(115200);
  FPS=0;
  fpsCounter = 0;
  fpsTime=millis();    
}
void loop() {
  
  printScreen();
}