An Energy-Efficient Deep Reinforcement Learning Accelerator With Transposable PE Array and Experience Compression

In this letter, we propose a low power deep reinforcement learning (DRL) SoC, supporting CNN and learning-optimized RNN, and fully connected layer. The adaptive reuse of weights and inputs, and data encoding/decoding techniques reduces power consumption and peak memory bandwidth of DRL processing by 31% and 41%, respectively. The 65-nm 16-mm
²
chip achieves a peak 2.16 TFLOPS/W at 0.73 V and 204 GFLOPS at 1.1 V with 16-bit data.