If you show your existing code, we could see what you mean by "feedback", and perhaps we can find how to improve it.

What about something like this? I think you can actually also do it without using the PImage as a texture. I just did that to get rotation.

The example you link to uses the pixels array and array copy and still gets good performance.

I suggest you describe what you are trying to do and if possible post some of your code.

I forgot that one used pixels. Here is another that doesn't:

void setup() {
  size(400,400);
  noStroke();
  background(0);
}

void draw() {
  ellipse(mouseX, mouseY, 20, 20);
  
  pushMatrix();
  translate(width*.5,height*.5);
  scale(1.75);
  rotate(frameCount*.01);
  tint(204,24);
  image(g.get(), -width*.5,-height*.5,width,height);
  popMatrix();
}

Well I managed to get 60fps (Processing caps it at that anyway). If I use framerate(1000) in setup I effectively remove the cap and I get 200-220 fps.

In this sketch the only thing that might take some time is the image command but this uses core Java classes which will have a fast implementation in the JVM.

I tried to do it with the pixels[] array but that turns my sketch into a 10 fps slide show.

If you have to manipulate individual pixels then you don't have a lot of choice but play with the pixel array. The fps will then be heavily dependant on the number pixels you access per frame (and your computer hardware)

This is just a thought and others might shoot me down, but you might look at using Processing V2 with the new OpenGL renderer.

Thanks for your help so far. I can't really test the performance right now because I'm rendering some 3d stuff in the background. But my resolution is 1280x720. So the performance already becomes a lot worse with that. I don't really need to manipulate the pixels. I just want to show the last frame a bit scaled and rotated to just get some classic video feedback.
In VVVV it's quite easy to do and it doesn't feel like  it's eating a lot of performance. So I guess this must be possible in Processing too.

How would I do it with OpenGL? I'm already using Processing 2.0.

So nobody can help me with this? Or maybe just give me a hint if FBOs are the right way to do this?
The resolution I have to deal with is now 1400x788. So the standard way is not an option anymore.

I have done some experiments with OPENGL vs JAVA2D in Processing 1.5.1 and 2.0a5 using the code pasted above with a resolution of 1280x720.

Framerates for       1.5.1       2.0a5
JAVA2D               39-44       8-9
OPENGL               9-12        18-21

I am not sure why the JAVA2D performance drops but the improved performance for the new OpenGL renderer is impressive.

As to FBO's, according to the Red Book FBO's should provide significant improvements but would involve making direct OpenGL method calls. Might be worth posting a query about OpenGL access and FBOs on the 'Integration and Hardware' forum.