Waveform compression: per-signal encoding barely beats FST/FSDB. The big ratios come from cross-signal redundancy or design-aware storage

Measured on a ~1.7 GB FSDB of a multi-core SoC sim (~6.3 G value changes across non-memory signals): - FSDB on disk is already ~0.27 B per value change. FST is in the same class. Don't assume these old formats are bloated. - The best per-signal adaptive codec (varint time deltas plus, per signal, XOR-delta, dictionary, byte-plane or period+exceptions, then zstd-3, block-indexed for random access) reached ~0.25 B/change, only ~1.1x smaller. - Hashing value streams and storing byte-identical ones once (port copies of wide buses) reached ~0.13 B/change, ~2.1x smaller. Wide 385-512-bit buses were 66% of the bytes. - A naive deflate-per-signal viewer cache was 11x LARGER than the FSDB.

Tools claiming 25-500x over FST (e.g. design-specific formats generated per RTL) almost certainly exploit the design. Internal signals are functions of inputs and state, so storing only what the RTL can't derive and re-simulating yields huge ratios. The catch: viewing needs the design plus an evaluator. An interpreted evaluator (~0.4 ms per eval) is far too slow for interactive windows, where 100 signals x 2000 cycles is ~200k evals. Zoomed-out views need every cycle evaluated, unless multi-resolution summaries are stored. That claws back size.

Cheaper middle ground with no RTL: detect relations in the dump itself, such as identical streams, time-shifted pipeline copies, inversions, and bit slices or concatenations.

Probe pitfall: a measurement probe that reloads each signal batch over many time slices, or tries ~40 codecs per signal in one process, makes a full pass take hours. Shard it across processes with RSS guards, and project from complete instances (identical tiles have near-identical stats) instead of waiting for all signals.