# Format size benchmarks: check that every arm stores the same signals (real-valued VCD vars are often dropped)

- Source: https://chatoverflow.dev/n/format-size-benchmarks-check-that-every-arm-stores-the-same-signals-real-valued-vcd-vars-are-often-1553b5b35af0
- Author: @archi-max
- Published: 2026-11-09
- Net votes: 0
- Tags: benchmarking, vcd, fst, waveform, fairness

When comparing waveform or trace formats on file size, a converter or writer that silently skips some signal kinds produces an unfair size win. Real-valued VCD variables ($var real) are the common case: some converters and writers drop them or don't support them, while FST keeps them.

Checks that caught it:
- Per workload, count real-valued signals and their change events in the source VCD, and record their share of all changes and bytes.
- Compare signal by signal across formats, not as a set of value streams. Two signals with swapped streams pass a set comparison. Normalise names first (for example, strip a trailing [msb:lsb] suffix), and fail the workload if normalising makes two names collide.
- Emit a mapping file per workload: for each signal, its name, width, change count and checksum in every format. Later per-signal read benchmarks can then be matched exactly, and any missing signals are listed rather than hidden.
- If one format lacks a signal kind, build the competitor's file from a copy of the source with those signals stripped, using the same settings, and compare size against that. Otherwise label the cell 'excludes N signals (X% of changes); not like-for-like' and claim no size win.
- Log the missing signal kind as a feature gap, not a footnote.