In-simulation waveform capture: per-signal callbacks are cheap for tens of signals and blow up for thousands
These are field results from building a custom waveform-capture library that hooks into a commercial Verilog simulator, compared with the simulator's own dumper.
Small signal sets (about 50 signals): a DPI tap at #0, which settles once per time step, with batching was roughly 2% simulation overhead. The vendor dumper was about the same. Keep the expensive work off the simulation thread: the call only appends (time, id, value) to a log, and a separate flusher thread compresses and writes. That cut overhead below the vendor dumper.
VPI value-change callbacks, which see every glitch, cost about 3x the vendor dumper's overhead.
Thousands of signals: tapping about 6K signals made the simulation about 3x slower (86 s against 29 s). Multi-dimensional memory arrays could not be tapped at all.
Practical policy: - Use custom in-simulation capture only for small selected signal sets. - For wide scope, let the simulator's native dumper write and convert afterwards. Conversion cost lands after the simulation, not inside it. - Measure overhead against a no-dump baseline on the same build and box, with shuffled repeats. Single runs mislead at the 1-2% level.