DPD Arena v6 APA B: full-window training on APA_200MHz_b

Arena compares DPD architectures and sizes through one fixed TRes-GRU PA per condition. The PA used for training is also the PA used for validation and test. GRU and GMP remain DPD contestants; there are no additional learned PA judges. The published Arena uses only APA_200MHz_b (200 MHz, 256-QAM), with one measured-source PA condition. Its 23 model entries and 233 test cases retain the original completed v6 observations and scores. Synthetic PA datasets and ILC-based entries do not enter Arena rankings. PA and DPD train, validation and test inputs are all unchanged samples from the original measured captures. Cascade outputs are predictions through the frozen PA, not new predistorted hardware measurements.

PA identification and selection

PA and DPD use the existing disjoint measured partitions. Train three freshly initialized H37 TRes-GRU candidates (4,871 real parameters), with seeds 0, 1 and 2. Each candidate receives 240 full passes over every 200-sample window at stride 1, shuffled without replacement, batch 64 including the final partial batch. AdamW uses initial learning rate 0.005 and weight decay 0.01. Minimum whole-validation-record NMSE selects the checkpoint. A plateau scheduler halves the rate after 10 unimproved epochs (0.01 dB absolute threshold), to a floor of 0.0001. Record the validation improvement over the last 30 epochs.

Older APA PAs trained on a different partition remain ineligible because they saw samples belonging to the current test set. All PA candidates and seed selections are frozen before loading any test arrays. Test NMSE is reported after selection and never changes the chosen PA. The validation residual spectrum is reported as an additional diagnostic.

Freeze one selected PA for all DPD architectures on each dataset. Rank displays its parameters, validation/test NMSE and checkpoint hash. PA cost is excluded from DPD FoM. See PA qualification.

Original measured inputs and partitions

Every saved x_train, x_val and x_test is an unchanged float32 slice of the original capture; y retains its paired measured PA output. No generated waveform, duplicated data, padding or resampling supplies missing samples. The independent audit compares every split with the original capture arrays and verifies nonoverlap, source bounds and file hashes. The protocol metadata is in calibration-v6-apa-b.json.

APA_200MHz_b contains independently timed LTE carriers. Its original 19,662-sample validation/test records are shorter than the 32,768-sample useful symbol, so they cannot support complete symbol EVM. Repartition the original capture before PA or DPD fitting.

Use only the original input to find CP correlations independently per carrier, with the existing fixed kurtosis timing refinement. Reserve the last complete symbol per carrier that leaves the fixed 200-sample end context. Test begins 200 samples before the earliest of those symbols. Split the preceding samples 80/20 into training and validation, with 200 unused samples at each boundary. No PA output or DPD outcome determines timing or boundaries.

Dataset	Train samples	Validation samples	Test samples	Complete scored test symbols
APA_200MHz_b	25,307	6,327	66,270	1 independently timed symbol per carrier, 32,768 samples each

APA bounds are [0,25307), [25507,31834) and [32034,98304). The fixed metric context is 200 original samples at both ends of the test record. Spectral metrics and the power gate use the middle interval; EVM uses complete symbols within it. Receiver channel isolation uses the full held-out record including its original context. No test data are borrowed from training or validation to complete an EVM symbol.

PA inference carries continuous recurrent state through each record, including internal chunks needed by cuDNN. PA qualification compares the full held-out measured output from zero initial state; DPD metrics use the predefined context.

Training and the common transmitter envelope

The sweep samples at most 250, 500, 1,000 and 2,000 real parameters. Each distinct configuration is trained once, stored at its smallest fitting budget, and can enter every larger budget ranking. Hidden-size models use their largest registered size within each sampling budget; polynomial orders remain fixed before scores are observed. Templates retain their 4,096-parameter admission limit. Missing configurations are unavailable, not zero-valued observations.

Each neural configuration receives seeds 0, 1 and 2 and 240 complete training passes, using every valid 200-sample window at stride 1. A seeded NumPy permutation chooses the order each epoch; all backbones share the same orders for a dataset and seed. Batch size is 64, and the final partial batch is never dropped. The order hash is SHA-256 of the concatenated epoch permutations as little-endian signed 64-bit integers.

Dataset	Windows per epoch	Updates per seed	Window exposures per seed
APA_200MHz_b	25,108	94,320	6,025,920

The cascade training loss remains MSE on the central 100 samples of each frame. AdamW uses learning rate 0.005, weight decay 0.01 and gradient clipping at 200. Validation occurs every epoch, on the original held-out validation input. Its spectral estimator uses 4,096-sample Hann/Welch segments, 50% overlap and a 200-sample context excluded at each end. Select the checkpoint with feasible validation output power first, then minimum 0.5 * (in-band error dB + max(ACLR_L, ACLR_R)). A tie keeps the earlier checkpoint. The scheduler uses this same spectral objective, factor 0.5, patience 10 epochs, 0.01 dB absolute threshold and minimum learning rate 0.0001. This validation in-band error is explicitly not demodulated EVM; only complete held-out test symbols supply the EVM entering final FoM.

Training workers load only train/validation arrays. All checkpoints in a submission are frozen before test inputs are loaded; the official matrix prefetches every configuration before beginning its test phase. Validation metrics never enter final FoM or rank. Training state is saved each epoch so interrupted local work can resume the same optimizer, scheduler, window order and RNG state, with no extra or omitted updates. Resume state is bound to the exact configuration, data, PA, training sources, seed and execution device.

Reviewed CUDA models replay forward/backward and clipping while AdamW remains eager. Zero-threshold Delta cells and BOJANET use algebraically equivalent dense GRU training operations; PGJANET, DVRJANET and APNRRU may compile their existing forward on CUDA. These paths retain FP32 parameters, gradients and optimizer state, with numerical parity tests. TF32 is disabled for DPD training and test on both CUDA hosts. Validation, exported weights, test inference and operation counts retain the original backbone. Optional seed concurrency uses private models, PAs, optimizers, permutations and CUDA streams inside one GPU process; all initialization and compilation finish before updates start.

The longer fixed budget addresses the v5 undertraining finding. A 240-epoch budget does not by itself prove every architecture is fully converged. These remain reproducible fixed-recipe results, rather than architecture optima or a fresh blind holdout after protocol development.

The gain target and radial PA-input peak limit come from the measured training partition. A common linear-gain baseline is calibrated using that training input only. Every seed and condition must keep output power within ±0.5 dB of the fixed target; output is never renormalized after prediction.

MP/GMP least-squares fits use indirect learning on the training capture only: regression inputs are PA(x_train)/G, and targets are x_train. They do not consume ILC-generated data. Their orders and regularization are fixed presets; validation/test samples do not fit coefficients. The gradient-trained GMP uses the neural train/validation workflow. Deterministic methods run once and do not receive invented seed variability.

ILC, including the former offline ILC→MP entry, is excluded from the Arena catalogue, submissions, rankings and Pareto fronts. The historical ILC→MP path used only the first 16,384 training samples and then deployed a fixed MP; it did not iterate ILC on test data. That distinction does not make it an Arena entry under this protocol. Studio's separate ILC workflow remains available. See the data-flow audit.

Offline DPD uses overlapping 200-sample windows and retains their middle 100 samples. Stateful variants reuse the base weights but carry their state and are evaluated and ranked separately. Noncausal context in offline models does not imply a causal deployment.

EVM and ACLR

For EVM, take each complete useful OFDM symbol, remove its CP, FFT it on the declared grid and keep only the published occupied bins. For APA, first frequency-shift and isolate each carrier independently, then use its frozen input-synchronized FFT start and all 1,200 occupied subcarriers (±600, excluding DC). This covers the full LTE 20 MHz allocation before the capture-rate doubling, not the older constellation preview’s central-600 subset. No per-subcarrier fit or output-dependent timing is used. Compare the PA output Y with the known reference R = Gx, removing one common complex gain:

c = sum(conj(R) Y) / sum(abs(R)^2)
EVM = sqrt(sum(abs(Y-cR)^2) / sum(abs(cR)^2))
EVM_dB = 20 log10(EVM); EVM_percent = 100 EVM

This is a reference-aided engineering EVM, not a complete LTE conformance receiver. The standard receiver reference is after CP removal and FFT (ETSI TS 136 104, Annex E). No full-record band-error fallback is allowed. Missing symbol metadata or a record without a complete symbol raises an error. In particular, Parseval's identity does not make the previous APA band-error proxy equivalent to symbol-demodulated EVM.

ACLR is measured from the cascade output itself, using the fixed opendpd-spectral-v2 Welch estimator and carrier bands:

ACLR_L/R = 10 log10(P_adjacent_L/R(y) / strongest in-band carrier power(y))
ACLR_worse = max(ACLR_L, ACLR_R)

The convention is negative dBc, lower is better. This is the OpenDPD carrier normalization, not a standards-conformance certification. Raw left/right, baseline and reference ACLR remain available. The adjacent-error ratio (AER) uses y-Gx; it remains a separately named diagnostic and never substitutes for output ACLR in selection, scoring or plots. NMSE is also diagnostic.

Figure of merit

For each configuration with actual DPD parameter count P and arithmetic count A, define the following relative to its common no-DPD baseline:

q_seed,condition = 0.5 (EVM_baseline_dB − EVM_DPD_dB)
                 + 0.5 (ACLR_worse_baseline − ACLR_worse_DPD)
q_seed = q_seed,condition  # one measured dataset per board
Q = mean_seed(q_seed)
FoM = Q − 5 log10(P/1000) − 5 log10(A/2000)

Every size uses fixed references of 1,000 parameters and 2,000 operations. Halving both costs with unchanged EVM and ACLR increases FoM by 3.0103 dB. The equal quality/cost weights are an explicit policy choice. FoM is not a physical RF efficiency, latency, energy or silicon-area measurement.

All final performance metrics, quality gates, FoM and ranking observations come from the held-out test split. Parameters and operation counts are properties of the frozen DPD; validation diagnostics never contribute to the final score.

Report sample standard deviation across seeds separately. The mean-minus-SD quantity remains a diagnostic, not the main score. Positive mean quality and the output-power envelope are required to rank; no 3 dB threshold applies. Retain negative FoM for otherwise valid configurations. Never turn missing sizes or invalid configurations into zero-score measurements.

The main table ranks individual configurations. Optional backbone summaries show their best observed configuration and are explicitly descriptive test summaries, not an unbiased estimate of a model selected without test results. Budget rankings admit every configuration whose actual size fits the cap.

Hardware-agnostic cost model

Only DPD arithmetic and real parameter count enter FoM. OPs = real MUL + ADD per output IQ sample in algorithmic steady-state execution, batch 1, dense FP32. This excludes PA cost, training cost, host timing and duplicated work from the overlap-window evaluation harness. Complex coefficients count as two real parameters and their conventional product as four real MULs.

The existing itemized arena_ops ledger and nonlinear reference table apply: most scalar nonlinear functions cost 1 MUL + 1 ADD using at most 512 uniform first-order segments with 2^-12 reference-domain absolute error; Hardswish costs 2 + 2 and atan2 3 + 4. Delays, indexing, signs, exact powers of two and table reads are free. No sparsity, quantization, pruning or fusion is credited. These assumptions are published so another implementation can use a different cost model. QGRU runs FP32; Delta thresholds are zero.

Pareto plots

Rank contains EVM vs parameters, ACLR vs parameters, EVM vs operations and ACLR vs operations. Horizontal cost axes are logarithmic; lower and left is better. Each point is an actual configuration, with per-seed mean and sample standard deviation. Each measured dataset has its own board and front. Never mix protocols, PA conditions, or offline/stateful execution when computing a front. Shipped and workspace results retain separate cohorts.

A point is dominated if another eligible point is no worse in both plotted coordinates and strictly better in at least one. Exact duplicate points remain nondominated. Each projection has its own front; a two-dimensional EVM front is not a joint four-dimensional EVM/ACLR/parameters/operations front. Grey crosses preserve unranked observations. Connecting lines are visual guides, not evaluated intermediate models.

Reproduction and integrity

benchmark/retrain_arena_pa.py refits PAs on the unchanged v4 measured partitions and writes calibration-v6-apa-b.json. The packaged original capture and measured partitions retain their hashes and boundary metadata. The historical benchmark/prepare_measured_arena.py utility requires its archived prior calibration; it is not needed to reproduce the published APA B matrix. The runner binds training sources, data, PA weights, metrics and frame draws to its training fingerprint. The protocol additionally binds scoring sources, presets and the arithmetic ledger. All DPD configurations are retrained under v6. Generated-input fits and earlier fixed-budget fits are not reused. Prior protocol score bundles do not enter v6 rankings.

Seeded fits may run on CPU or CUDA, including different devices across seeds. Every fit retains the prescribed budget and frame draws and records its actual training device. Completed checkpoints can be reused across devices after source and weight-hash verification. CPU and CUDA kernels can round differently; numerical replay therefore retains the recorded PA evaluation device while changing the DPD chunk size. Arithmetic costs follow the same published ledger on every device.

The independent auditor recomputes quality, costs and ranks without the production aggregation; validates complete cases, budgets and checkpoint selection; verifies operation counts; and replays stateful DPD with different chunk sizes plus a second EVM implementation. See the reproduction guide and current results.

Release scope and provenance

The dpd-arena-v6-apa-b release contains only APA_200MHz_b results. All 218 independent fits were completed and frozen before the original test evaluation. benchmark.focus_arena_reference verifies the archived matrix, copies its sealed APA B checkpoints, requires every raw observation and score to remain identical, and reruns the independent audit. There is no retraining or test-based checkpoint selection during projection. The calibration-v6-apa-b.json file contains only APA B. Narrowing calibration changes the training identity, so the projection records both the original and release identities while retaining identical weights, training history and RF observations. Other datasets are outside the public Arena.