OpenDPDv2: Unified Framework for Neural Network Digital Predistortion
OpenDPDv2, a newly developed open-source framework, integrates power amplifier (PA) modeling, neural network-based digital predistortion (NN-DPD) learning, and optimization tailored for deployment. It features TRes-DeltaGRU, a delta-RNN design that incorporates a lightweight temporal residual path, which facilitates effective performance even with high temporal sparsity and allows for joint optimization with fixed-point quantization. In experiments conducted with a 3.5 GHz GaN Doherty PA using a TM3.1a 200 MHz 256-QAM OFDM signal, the FP32 model recorded an Adjacent Channel Power Ratio (ACPR) of -59.9 dBc and an Error Vector Magnitude (EVM) of -42.1 dB. When employing 56% temporal sparsity and W12A12 quantization, the model maintained -51.8 dBc ACPR with only 450 active parameters. The framework targets the reduction of digital back-end complexity in wideband RF PAs, tackling a significant obstacle in NN-DPD implementation.
Key facts
- OpenDPDv2 is an open-source end-to-end framework for NN-DPD.
- It unifies PA modeling, NN-DPD learning, and deployment-oriented optimization.
- Introduces TRes-DeltaGRU, a delta-RNN architecture with a temporal residual path.
- Supports joint optimization with fixed-point quantization.
- Tested on a 3.5 GHz GaN Doherty PA with a TM3.1a 200 MHz 256-QAM OFDM signal.
- FP32 model achieves -59.9 dBc ACPR and -42.1 dB EVM.
- With 56% temporal sparsity and W12A12 quantization, uses 450 active parameters.
- Maintains -51.8 dBc ACPR under quantization and sparsity.
- Targets reducing digital back-end complexity for wideband RF PAs.
Entities
—