The lack of improvement that I reported with #549 has always bothered me; it was an impression, not a measurement. So, I went back to it, and tried timing plt.draw() with the array of subplots. I find there is an improvement on my machine, from 0.53 s per draw to 0.37 s per draw; it just wasn't dramatic enough to be visually obvious when I was testing by manual zooming and panning. Maybe it is still worthwhile, if on a substantial sample of machines the improvement is more tangible. The extra complexity does make me hesitate, though.

I tend to agree -- anytime caching is introduced to speed something up, things can get much more brittle to change and full of unintended side effects. #549 does somewhat illustrate where a good 40% of the time is going though -- there just might be a better way to speed it up than just caching.

Closing this as no work has been done on it in 2 years and the assumption that Moore's Law is on our side.

Is there any update on this? I currently need to have a figure containing 16 x 33 subplots and it takes ages to load.

It sounds like the relevant issue is #6664, in particular #6664 (comment).