Question

I am wondering why post-layout simulations for digital designs take a long time?

Why can’t software just figure out a chip’s timing and model the behavior with a program that creates delays with sleep() or something? My guess is that sleep() isn’t accurate enough to model hardware, but I’m not sure.

So, what is it actually doing that takes so long?

Thanks!

Answer 1

1. Post layout simulations (in fact – anything post synthesis) will be simulating gates rather than RTL, and there’s a lot of gates.

2. I think you’ve got your understanding of how a simulator works a little confused. I say that because a call like sleep() is related to waiting for time as measured by the clock on the wall, not the simulator time counter. Simulator time advances however quickly the simulator runs.

A simulator is a loop that evaluates the system state. Each iteration of the loop is a ‘time slice’ e.g. what the state of the system is at time 100ns. It only advances from one time slice to the next when all the signals in it have reached a steady state.

In an RTL or untimed gate simulation, most evaluation of signals happens in ‘zero time’, which is to say that the effect of evaluating an assignment happens in the same time slice. The one exception tends to be the clock, which is defined to change at a certain time and it causes registers to fire, which causes them to change their output, which causes processes, modules, assignments which have inputs from registers to re-evaluate, which causes other signals to change, which causes other processes to re-evaluate, etc, etc, etc…. until everything has settled down, and we can move to the next clock edge.

In a post layout simulation, with back-annotated timing, every gate in the system has a time from input to output associated with it. This means nothing happens in ‘zero time’ any more. The simulator now has put the effect of every assignment on a list saying ‘signal b will change to 1 at time 102.35ns’. Every gate has different timing. Every input on every gate will have different timing to the output. This means that a back-annotated simulation has to evaluate lots and lots of time slices as signals are changing state at lot’s of different times. Not just when the clock changes. There probably isn’t much happening in each slice, but there’s lots of them.

…and I’ve only talked about adding gate timing. Add wire timing and things get even more complex.

Basically there’s a whole lot more to worry about, and so the sims get slower.