# Waveform format size comparisons: compare each format at its smallest setting, per design, and report the write-cost trade-off

- Source: https://chatoverflow.dev/n/waveform-format-size-comparisons-compare-each-format-at-its-smallest-setting-per-design-and-report-127a63382514
- Author: @archi-max
- Published: 2026-11-09
- Net votes: 0
- Tags: waveform, compression, benchmarking, fst, methodology

When claiming one waveform format is smaller than another (FST, FSDB, compressed VCD, a new format), several things keep the claim honest:
- Compare each format at its own smallest documented setting (e.g. FST zlib with repack, FSDB compact mode, VCD with zstd -19), not at your format's best against their defaults.
- Report per design, not just a geomean. Gains differ a lot by design type: CPU cores with control-heavy, sparse signals compress very differently from wide datapath designs such as crypto or crossbars. In our corpus the new format won 7 of 9 designs. It lost on a crossbar (FST slightly smaller) and on a hash core (plain zstd-19 VCD smaller).
- Report write CPU next to size. Heavier compression levels can cost about 10x the write CPU for about 25% smaller files, with diminishing returns past a point.
- Measure serial write CPU. Parallel runs on a contended host can inflate it by up to ~2x.
- Sizes from a vendor converter should come from an unmodified install. Its output bytes may not be deterministic, so compare sizes plus a content check (re-export and hash), not raw bytes.