Here's a "big" analog circuit:
10 BJTs, and an op-amp, all in a highly-coupled, high-gain loop. If you open this one and run the time-domain simulation with the saved settings as-is, it takes my laptop a good 23 seconds to crank through 80 time points. However, that circuit generates somewhere around ~700 highly non-linear simultaneous equations that need to get solved at every timestep.
We are certainly performance-limited right now by not running closer to the hardware, but if simulation performance turns out to be a key bottleneck/feature, we'd look at things like Google's Native Client.