DPD linearization for wideband PA - memory effects handling?

Wideband DPD lives or dies on memory effects. What memory depth and model do you actually run, and how much ACPR improvement do you get?

Wideband DPD that "works on CW two-tone" and fails on real traffic is almost always memory / thermal / bias dynamics, not a missing polynomial term.

What usually moves the needle:
1. Memory depth and basis that match the PA's envelope bandwidth and thermal time constants - short Volterra / GMP memory for electrical memory, plus slower terms or a separate thermal model for long tails.
2. Training signal that looks like traffic (PAPR, bandwidth, duty). Training on narrowband or lightly clipped waveforms under-learns the memory you see in the field.
3. Capture length and alignment: insufficient observation of post-burst recovery makes residual ACLR look "random" when it is actually settling.

Useful report: signal BW / PAPR, memory depth you settled on, ACLR before/after, and whether residual tracked temperature or just envelope. Share if you closed the loop with observation RX bandwidth ≥ PA envelope - undersampled observation is a common silent fail.

— rfforum 编辑部 / Editorial Team

How's the power consumption on this? Our project has a tight power budget.