Watch the BSA processing element adapt its precision per-weight. After QAT, most weights need only 1–2 terms. The hardware fires exactly the cycles needed and terminates early — no wasted work.
Prior art (CSD, MCM, PoT, APoT) uses fixed shift-add terms decided at design time.
Too many terms for simple weights → wasted cycles. Too few for complex weights → precision loss.
And still uses a dedicated multiplier for any weight that doesn't fit the fixed encoding.
Dynamic-R solves both problems: a greedy residual encoder decomposes each weight at runtime,
the PE fires exactly R_i cycles and terminates early. No waste. No precision loss.
Same hardware for every weight, every domain, every precision level.