DPD for wideband 5G - memory polynomial order vs performance

400MHz instantaneous bandwidth 5G NR signal, memory effects are brutal. Going beyond 7th order MP barely helps but doubles the coefficient count. What's the sweet spot you guys use for wideband DPD?

For ~400 MHz IBW 5G NR, memory-polynomial order is rarely the only knob — past ~5th–7th nonlinear order you often get coefficient bloat with little ACLR/EVM return unless the memory depth and basis match the PA’s dynamics.

What tends to work better than “just go to 9th/11th”:
1. Cap nonlinear order around where the measured AM/AM–AM/PM and ACLR stop improving on a sweep; put budget into memory depth / taps and orthogonalization (or GPR/DDR variants) instead.
2. Band-limited or decomposed bases (e.g. GMP / DDR / piecewise) so wideband memory is modeled without a full dense MP of huge order.
3. Indirect learning vs direct learning: direct often needs fewer “useless” coeffs once you have a stable observation path; still validate on a held-out power / bandwidth corner.
4. Crest-factor and supply effects: if envelope tracking or a soft supply is in play, a pure MP of any order will look stuck — fix the observation model first.

Sweet spot we hear most for ~100–400 MHz class NR PA DPD in production is modest order (often ≤7) with carefully chosen memory depth, not a 11th-order monster. What PA technology (GaN/LDMOS/CMOS) and are you optimizing ACLR, EVM, or both under a fixed LUT size?

— rfforum 编辑部 / Editorial Team