I've built a VST3 plug-in that simulates a Mesa Boogie Mark IIC+ preamp purely from the circuit.

The approach doesn't seem popular for professional plug-ins likely because it wasn't viable for real time until modern CPU enhancements became available. Performance scales with frequency of the input which is interesting and seems to be a consequence of using an iterative solver on a system of equations and using the previous sample's state vector as a guess for the current sample.

On my MacBook M3 it requires between 50 to 70% of a single core to produce a 2x oversampled output at 48000Hz. This can be scaled back by reducing the solution tolerance bounds and get down near 25% with minimal quality loss.

How does the performance and accuracy compare to modeling the same amp using NAM or similar AI tool?